IntroductionNeonatal pulmonary diseases are among the leading causes of morbidity and mortality in newborns. These conditions, which can be congenital or acquired, pose a serious threat to neonatal health and require prompt diagnosis and treatment. Clinical diagnosis typically relies on chest X-rays, blood gas analysis, and other physiological indicators. However, manual interpretation of neonatal chest X-rays is time-consuming, subjective, and highly dependent on the experience of pediatric radiologists. In recent years, deep learning has emerged as a promising tool for automating medical image analysis and assisting clinical decision-making1, 2, 3. Deep learning has also shown remarkable success across various biomedical imaging applications, including breast cancer4, diabetic retinopathy5, anti-cancer drug response prediction6, and brain tumor diagnosis7, 8.Despite these advances, most existing public datasets (such as CheXpert, NIH ChestX-ray14) are designed for adult populations, while neonatal lung diseases present distinct pathological and radiographic patterns. The absence of a dedicated neonatal pulmonary dataset has become a critical bottleneck limiting the development of reliable AI-assisted diagnosis in neonatal care. To address this gap, we constructed the first neonatal pulmonary ailment dataset (NPA), which contains two subsets targeting distinct diseases-neonatal respiratory distress syndrome (five-class classification) and neonatal aspiration syndrome (binary classification). The NPA dataset includes 828 normal cases, 683 cases of respiratory distress syndrome, and 886 cases of aspiration syndrome, establishing a new foundation for intelligent neonatal lung diagnosis.However, a key challenge arises from the inherent class imbalance in clinical neonatal data, where healthy cases significantly outnumber diseased ones. This long-tail distribution causes deep learning models to overfit the dominant class, leading to degraded sensitivity toward minority (diseased) categories. Classical supervised learning depends heavily on labeled data, which is costly and time-consuming to obtain in medical settings. Unsupervised learning and consistency-based methods can help mitigate this problem by exploiting unlabeled data9, 10, 11, but they do not directly solve the imbalance issue within the labeled portion.Researchers addressing the scarcity and imbalance of neonatal lung disease data commonly rely on data augmentation techniques such as CutMix12, Cutout13, and Mixup14. These methods have proven highly effective in natural image classification by enriching training diversity and improving model generalization. However, in medical imaging, their random pairing strategy can be problematic: mixing two diseased cases with different lesion types may distort pathological semantics, while random cropping may even remove critical lesion regions. Consequently, conventional augmentation not only fails to guarantee meaningful disease synthesis but can also confuse the model’s discrimination of disease-specific patterns, especially under severe class imbalance. In short, uncontrolled randomness introduces uncertainty that hinders reliable representation learning. Despite the growing use of augmentation in medical image analysis, little attention has been given to controlling semantic distortion and uncertainty during synthetic sample generation—creating a key research gap this work aims to address.To overcome these limitations, we propose a novel Polluted CutMix framework. Unlike traditional CutMix, which randomly selects any two images for blending, our approach constrains the pairing to one diseased and one normal case. This targeted combination ensures that the generated images always contain clear disease-related semantics, while simultaneously increasing the number of effective diseased samples and mitigating the long-tail distribution. The cropping and pasting regions remain random to preserve diversity; however, this randomness may still occasionally exclude or distort lesion areas, introducing uncertainty into the augmented data.Recent works have further emphasized the importance of uncertainty modeling and reliability in medical AI15, 16, demonstrating that accurate estimation of prediction confidence can significantly enhance diagnostic stability. Inspired by these findings, we incorporate an uncertainty-aware module into our framework to promote more reliable decision-making. Inspired by these studies, we further integrate an uncertainty-aware module into our framework. By quantifying prediction uncertainty through Monte Carlo sampling1718, the framework automatically identifies and filters out unreliable pseudo-labels that arise from ambiguous or noisy regions. This mechanism prevents error propagation and ensures that only confident samples contribute to learning.For uncertainty filtering, we further employ a student-teacher self-ensembling framework19. The teacher network, updated as an exponential moving average (EMA) of the student’s parameters, produces stable pseudo-labels for unlabeled or mixed data. The student is trained to maintain consistency with the teacher’s predictions under perturbations, but only for low-uncertainty samples identified by the Monte Carlo estimator. Through the collaboration of Polluted CutMix and uncertainty-aware learning, the proposed framework effectively handles the challenges of medical data imbalance and unreliability, improving both generalization and diagnostic robustness.We conducted a comprehensive study on the effectiveness of each proposed component. Our contributions can be summarized as follows:1.We collected the country’s first up-to-date neonatal lung disease dataset, NPA, which includes two datasets: neonatal respiratory distress syndrome and neonatal aspiration syndrome.2.We find that sampling between different types of illness cases interferes with the model’s judgment, and we propose the Polluted CutMix framework to address this issue.3.We design an uncertainty-aware model that introduces an unsupervised loss function and performs uncertainty detection to enhance the model’s learning depth and generalization ability, effectively handling extreme cases caused by random cuts.4.We integrate the above two frameworks and conduct extensive testing on the existing dataset, demonstrating their effectiveness.Related workData augmentationData augmentation is a technique that increases the diversity of training data by applying a series of random transformations and expansions to the original data. Data augmentation can help models generalize more effectively to previously unseen data, thereby enhancing the model’s robustness and performance. Our Polluted CutMix is also a data augmentation technique that can effectively address the long-tail problem caused by an excess of single-type samples during the data preprocessing stage. Next, we will introduce several data augmentation techniques.Traditional data augmentation techniques, such as random rotation, random scaling, and random flipping20, typically alter the appearance of the input data to increase the model’s invariance and generalization capabilities. However, these traditional data augmentation methods do not change the labels and are not very effective in addressing the class imbalance problem caused by an excess of single-type samples.Recent studies have explored active learning and attention-based strategies to improve data diversity and model generalization21, 22. In contrast, our Polluted CutMix focuses on label-consistent image composition to address class imbalance and the overrepresentation of single-type samples. Moreover, Singh et al.23 and Banerjee et al.24, 25 proposed attention-guided deep learning frameworks that enhance feature extraction and emphasize disease-relevant regions in breast, lung, and thyroid imaging tasks. These attention-based methods illustrate how targeted feature emphasis can improve discriminative learning and model interpretability. Our Polluted CutMix can obtain new labels through preprocessing, thus solving the class imbalance issue.Region dropout randomly sets certain regions of the model to zero during training, similar to traditional node-level dropout, but it operates on the spatial dimensions of the feature maps. This helps reduce the model’s dependency on specific regions and enhances the model’s robustness to input variations. Methods for removing random regions from images2627 have been proposed to improve the generalization performance of CNNs. Target localization methods2829 also employ region dropout techniques to enhance the localization ability of CNNs. Polluted CutMix is similar to these methods, with the key difference being that the removed region is filled with a patch from another training image. DropBlock30 extends region dropout to feature space, demonstrating improved generalization. Polluted CutMix can also be performed in feature space, as shown in our experiments.Mixup31 and Polluted CutMix share similarities because both combine two samples, where the linear interpolation of the one-hot labels gives the true label of the new sample. Mixup samples exhibit local blurriness and unnaturalness, which confuses the model, particularly in localization tasks. Previously, variants of Mixup have been proposed32, 33; they perform feature-level interpolation and other types of transformations. However, these works often lack an in-depth analysis of localization capability and transfer learning performance. In contrast, Polluted CutMix demonstrates excellent transfer learning performance, and we validate the benefits of Polluted CutMix for image classification tasks.The traditional CutMix is inspired by the combination of Cutout and Mixup. It creates new training samples by cutting and mixing parts of two images. This method aims to improve the model’s robustness to partially occluded image regions, thereby enhancing the model’s generalization ability. However, existing CutMix typically randomly selects two images, which are then processed. In this scenario, for multi-classification problems, the effectiveness of randomly selecting images for cutting is poor. The presence of shadow regions is a crucial criterion in medical image classification. If two randomly selected images belong to different disease classes, randomly cropping a bounding box may damage important shadow regions, causing misclassification and degrading performance. Our Polluted CutMix addresses this issue. By processing images in fixed pairs—always combining a diseased case with a normal case by pasting lesion regions onto the normal image—we eliminate the damage to shadow regions that occurs when two diseased cases are mixed.Fig. 1Full size imageThe figure illustrates two modules: Polluted CutMix and Uncertainty-Aware Framework. After being contaminated with CutMix, the input image is fed into the teacher model as an unsupervised loss. Monte Carlo sampling is used to estimate uncertainty and determine which samples to use to guide student model learning.Uncertainty-aware self-ensemble model guided consistency learning frameworkTo enhance the model’s learning depth and generalization ability, we often introduce unsupervised losses. Unsupervised loss is a loss function computed from the model’s own predictions when no labels are available. Its introduction helps the model learn latent data representations, improves performance on unlabeled data, and reduces reliance on annotations, thereby boosting robustness and generalization. Common unsupervised losses include the reconstruction loss of autoencoders34, the generative loss of Generative Adversarial Networks (GANs)35, contrastive loss36, 37, and others.The unsupervised loss we introduce is a form of consistency regularization, which is a form of label propagation. It is derived from regularization techniques. Training samples that are similar to each other are more likely to belong to the same class. Therefore, the outputs should remain unchanged after receiving perturbations38, 11.To explicitly implement this consistency principle, we adopt a student-teacher self-ensembling framework19. In this setting, the student network serves as the main model, while the teacher network—updated as an exponential moving average (EMA) of the student’s parameters—generates soft pseudo-labels for unlabeled data. Both networks receive the same perturbed inputs, and the unsupervised consistency loss encourages the student’s predictions to align with the teacher’s outputs. This design stabilizes training, propagates reliable supervision to unlabeled samples, and effectively captures the intrinsic data structure even in the absence of ground-truth annotations.Compared to conventional consistency regularization, we choose relatively specific predictions for consistency learning, effectively mining more reliable information from unlabeled data. Instead of applying consistency regularization to all images, we introduce an uncertainty-aware self-ensemble model. Similarly, Banerjee et al.39, 40, 41 investigated uncertainty estimation, adaptive attention, and explainable deep models across diverse clinical domains, highlighting the critical role of prediction confidence in ensuring safe and interpretable medical AI systems.Uncertainty perception refers to the process in deep learning where the model perceives and utilizes the uncertainty of its predictions. When deep learning models make predictions on previously unseen data, they often exhibit a certain level of uncertainty regarding the confidence in their predictions. The goal of uncertainty perception is to effectively identify and leverage this uncertainty to improve the model’s generalization ability, robustness, and reliability. In deep learning, leveraging uncertainty perception can help improve the performance and robustness of models. Common methods include ensemble learning42, 43, Monte Carlo inference44, 45, and others.Self-ensembling strategy is a training method used to enhance the generalization ability and robustness of deep neural networks. It improves model performance by reducing uncertainty through multiple forward passes of the same sample during training and averaging the prediction results. Similarly, ensemble and efficient learning frameworks have been employed in prior works to enhance model stability and robustness46, 47. Specifically, their ensemble-based chest X-ray classification improved diagnostic reliability, while their efficient deep anomaly detection network achieved a balance between performance and computational cost. Different from these classification-oriented ensemble approaches, our uncertainty-aware self-ensembling framework explicitly models prediction confidence to guide semi-supervised consistency learning. The core idea of the self-ensembling strategy is to utilize the stability and consistency of the model itself to improve its generalization ability, reducing model uncertainty through multiple forward passes and averaging prediction results to enhance classification accuracy.Our method utilizes an uncertainty-aware self-ensembling model to guide the consistency learning framework. We perform multiple forward passes and average the prediction results, adding appropriate perturbations at each forward pass to obtain a dynamic threshold. Compared to conventional fixed thresholds, this dynamic threshold can more effectively adapt to the characteristics of the data and the model’s performance, thereby improving the model’s adaptability and performance. Only data with uncertainty below the threshold (high confidence) will be applied, determined by a mask.MethodOur Uncertainty-Aware Polluted CutMix framework aims to generate a sufficient number of high-quality, disease-labeled samples to augment the dataset and alleviate the long-tail issue caused by an overabundance of healthy cases.The overall framework is illustrated in Fig. 1. It consists of two main components: the Polluted CutMix module and the Uncertainty-Aware teacher-student framework. In the first stage, the Polluted CutMix module constructs mixed samples by combining diseased and normal chest X-rays. This operation increases sample diversity and balances the dataset by introducing controlled noise at the image level. The resulting Polluted CutMix images serve as unlabeled inputs for the unsupervised branch. In the second stage, these mixed images are processed by the teacher-student framework. The teacher model, updated through an exponential moving average (EMA) of the student’s parameters, performs Monte Carlo dropout to estimate prediction uncertainty. Based on the entropy-based uncertainty measure, a dynamic threshold is applied to determine which samples are reliable enough to guide the student model. Samples with uncertainty below the threshold contribute to the unsupervised consistency loss, while high-uncertainty samples are excluded to prevent noise accumulation.Through this interaction, the teacher model provides stable pseudo-labels for confident regions, and the student model learns both from labeled data (supervised branch) and reliable unlabeled data (unsupervised branch). This unified framework combines data augmentation and uncertainty-guided consistency learning to achieve robust performance under class imbalance and label scarcity.The following Sections will detail, respectively, the implementation of Polluted CutMix, the Uncertainty-Aware framework, the integrated UA-Polluted-CutMix framework, and the Neonatal Pulmonary Ailment dataset.Polluted CutMixOur Polluted CutMix is an improvement on CutMix, so we first introduce CutMix. CutMix is a data augmentation technique particularly well-suited for image classification tasks. It combines the concepts of image cropping (Cutout) and image mixing (Mixup). By cutting out a part of one image and pasting it onto another image, CutMix can create new training samples. CutMix aims to enhance a model’s generalization and robustness, especially when training data is limited or the data distribution is imbalanced. The principle behind our Polluted CutMix is similar to that of CutMix. In contrast to standard CutMix, which randomly selects any two samples from each mini-batch regardless of their classes, our Polluted CutMix enforces that one sample must be a healthy case and the other must be a diseased case, and pastes the diseased region onto the healthy sample. By directly “injecting” lesion regions into a large number of healthy samples, we can more precisely control data balance, retain the true appearance of lesions in the images, and avoid potential interference caused by randomly overlaying two different disease types.Here, \(x \in {\mathbb {R}}^{W \times H }\) and \(y\) denote a training image and its one-hot label, respectively. We first sample a ‘normal’ example \((x^{A}, y^{A})\) and a ‘diseased’ example \((x^{B}, y^{B})\) from the training set. A mixing coefficient \(\lambda\) is drawn from the Beta distribution Beta(\(\alpha\), \(\alpha\)); in our experiments we set \(\alpha\) = 1, so \(\lambda \sim \textrm{Uniform}(0,\,1)\,\). To generate a new sample \(({\tilde{x}},\,{\tilde{y}})\), we choose a rectangular crop region \(B\subseteq [0,W)\times [0,H)\) whose area satisfies$$\begin{aligned} \begin{aligned} |B|\;=\; r^{w}\times r^{h} \;=\; (1 - \lambda )\,W\,H\, \end{aligned} \end{aligned}$$(1)meaning that a fraction \((1-\lambda )\) of the pixels will be taken from \(x^{B}\) and the remaining fraction \(\lambda\) from \(x^{A}\). Concretely, we sample an aspect-ratio factor \(\rho \sim \textrm{Uniform}(0.5,\,2)\) and then set$$\begin{aligned} \begin{aligned} r^{w}&\;=\;\bigl \lfloor \sqrt{\, (1 - \lambda )\,W\,H \;\times \;\rho \,}\bigr \rfloor ,\qquad \\ r^{h}&\;=\;\bigl \lfloor \sqrt{\, \tfrac{(1 - \lambda )\,W\,H}{\rho } \,}\bigr \rfloor , \end{aligned} \end{aligned}$$(2)so that \(r^{w}\times r^{h}\approx (1 - \lambda )\,W\,H\). Next, we uniformly sample the top-left corner coordinates.$$\begin{aligned} \begin{aligned}&r^{x}\sim \textrm{Uniform}\bigl (0,\,W - r^{w}\bigr ),\qquad \\&r^{y}\sim \textrm{Uniform}\bigl (0,\,H - r^{h}\bigr )\,, \end{aligned} \end{aligned}$$(3)Thus, \(B = \bigl (r^{x},\,r^{y},\,r^{w},\,r^{h}\bigr )\) ndicates the patch on \(x^{A}\) to be removed and replaced by the patch cropped from the same coordinates of \(x^{B}\). We then construct a binary mask \(M\in \{0,1\}^{W\times H}\) by setting$$\begin{aligned} \begin{aligned} M_{i,j} = {\left\{ \begin{array}{ll} 0, & (i,j)\in B,\\ 1, & \text {otherwise}, \end{array}\right. } \end{aligned} \end{aligned}$$(4)so that pixels inside \(B\) are taken from \(x^{B}\), and pixels outside \(B\) remain from \(x^{A}\). Finally, the mixed image is formed as$$\begin{aligned} \begin{aligned} {\tilde{x}}&\;=\; M \odot x^{A} \;+\; (1 - M)\odot x^{B}, \\ {\tilde{y}}&\;=\; \lambda \,y^{A} \;+\; (1 - \lambda )\,y^{B}, \end{aligned} \end{aligned}$$(5)where \(\odot\) denotes element-wise multiplication. By enforcing that one sample is always ‘normal’ and the other is always ‘diseased’, our Polluted CutMix injects true lesion patches into healthy images, thus increasing the effective number of diseased examples and mitigating long-tail effects. Since the rectangular masks and Beta-sampling steps follow standard data-augmentation techniques, the computational overhead is minimal, allowing seamless integration with any backbone network. Importantly, these augmentations sit entirely outside the DenseNet architecture—our backbone remains unchanged—yet they enrich its input distribution. Unlike vanilla DenseNet, which relies solely on standard augmentations, Polluted CutMix specifically targets the minority “diseased” class by synthesizing realistic lesion-containing images, thereby complementing DenseNet’s feature reuse without altering its convolutional or connectivity layers.Uncertainty-aware frameworkRandom cropping and mixing strategies, while beneficial for data augmentation, may inadvertently select regions that contain little or no pathological information. In chest radiographs of diseased neonates, certain random crops might include healthy tissues or non-disease areas, introducing label noise and local ambiguity. If trained directly on such regions, the model may learn spurious correlations or uncertain features. To alleviate this problem and stabilize learning under partially informative supervision, we introduce an uncertainty-aware consistency framework that explicitly quantifies prediction confidence and suppresses the contribution of ambiguous regions during training. To enhance the model’s learning depth and generalization ability-while avoiding extreme cases where CutMix operations randomly select bounding boxes that eliminate important shadow regions and degrade performance—we introduce an unsupervised learning scheme guided by uncertainty. The source of this unsupervised signal is the set of images processed by the Polluted CutMix strategy. To exploit these unlabeled samples, we adopt a teacher-student framework. The student model (main network) is trained conventionally, whereas the teacher model (auxiliary network) provides stable targets for consistency training. Specifically, we use a DenseNet variant as the student model, and the teacher model is constructed as an Exponential Moving Average (EMA) of the student.At each training step t, the teacher’s parameters \(\theta '_t\) are updated based on the student’s parameters \(\theta _t\) and the previous teacher state \(\theta '_{t-1}\), following the EMA update rule:$$\begin{aligned} \begin{aligned} \theta '_t = \alpha \theta '_{t-1} + (1 - \alpha ) \theta _t \end{aligned} \end{aligned}$$(6)where \(\alpha \in [0, 1)\) is the momentum coefficient that controls the update rate. This mechanism allows the teacher model to accumulate knowledge over time from the student’s trajectory, effectively smoothing out training noise and providing more reliable targets for the unsupervised consistency loss.To mitigate the impact of extreme cases and prevent the teacher model from becoming unreliable due to noisy data, we employ an uncertainty-aware consistency training strategy. This mechanism enables the student model to selectively learn from more trustworthy targets, thereby improving the robustness and stability of the training process. Specifically, we apply the proposed Polluted CutMix to the training images and feed the resulting samples into both the student and teacher models. The teacher network serves two roles: generating soft pseudo-labels and estimating prediction uncertainty. Inspired by Bayesian uncertainty modeling, we estimate prediction uncertainty using Monte Carlo Dropout (MC Dropout). For each input sample, we perform T stochastic forward passes through the teacher model with dropout enabled and inject Gaussian noise to generate diverse predictions. This yields a set of softmax probability vectors \(\{p_t\}_{t=1}^T\), where \(p_t^c\) represents the probability of class c at the t-th pass. To quantify uncertainty, we use predictive entropy as a confidence measure, which has a bounded and interpretable scale. The entropy \(H({\hat{p}})\) for a given sample is defined as:$$\begin{aligned} H({\hat{p}}) = - \sum _{c=1}^C \left( \frac{1}{T} \sum _{t=1}^T p_t^c \right) \log \left( \frac{1}{T} \sum _{t=1}^T p_t^c + \epsilon \right) \end{aligned}$$(7)where C is the number of classes, and \(\epsilon\) is a small constant to avoid numerical instability. This entropy reflects the consistency of the model’s predictions across multiple passes. A lower entropy indicates higher confidence (i.e., lower uncertainty), while a higher entropy suggests that the model is uncertain about its prediction.Specifically, we define the uncertainty-aware consistency loss \(L_u\) to supervise the student model using reliable targets from the teacher. This loss computes the mean squared error (MSE) between the student model’s softmax prediction \(p_v\) and the teacher model’s prediction \(q_v\) for unlabeled data, but only for samples with low predictive uncertainty. To achieve this, we first compute the uncertainty estimate \(u_v\) for each sample v using the entropy-based criterion described in Eq. (7). A dynamic threshold H(t), which decreases over time, is used to gradually include more samples as training progresses. The consistency loss \(L_u\) is formally defined as:$$\begin{aligned} \begin{aligned} L_u&= \frac{1}{N_M} \sum _{v=1}^{N} \mathbb {1}_{\{ u_v< H(t) \}} \cdot \left\| p_v - q_v \right\| ^2 \\ N_M&= \sum _{v=1}^{N} \mathbb {1}_{\{ u_v < H(t) \}} \end{aligned} \end{aligned}$$(8)Here, N is the number of unlabeled samples in the mini-batch, \(N_M\) is the number of samples satisfying the uncertainty threshold, \(p_v\) is the softmax output from the student model, \(q_v\) is the output from the teacher model (with exponential moving average weights), and \(\mathbb {1}_{\{ \cdot \}}\) is the indicator function. This selective strategy ensures that only trustworthy pseudo-labels contribute to the unsupervised loss, thus reducing the impact of noisy targets and improving training robustness.This uncertainty-aware mechanism is applied externally to the DenseNet backbone, which remains structurally unchanged. While vanilla DenseNet optimizes only a cross-entropy loss on each mini-batch, our approach wraps the backbone in a teacher-student framework: the teacher’s exponential moving average weights provide soft targets, and Monte Carlo-estimated predictive entropy filters out unreliable pseudo-labels. Consequently, we enhance DenseNet’s learning dynamics under noisy or imbalanced conditions without modifying its convolutional or connectivity layers.UA-polluted-CutMixIn summary, we proposed two structures: the Polluted CutMix and the Uncertainty-aware Framework. We integrated the overall framework of these two methods. We utilized DenseNet48 as our network backbone. Then, the overall objective function is a combination of the supervised loss \(L_s\) and the unsupervised loss \(L_u\), as shown below:$$\begin{aligned} L = L_s + L_u \end{aligned}$$(9)where \(L_s\) consists of the weighted average of the cross-entropy loss and the Dice loss, with a weight coefficient of 0.5 used to evaluate the quality of the output when labeled data is input into the network. \(L_u\) computes the mean squared difference between the softmax outputs of the model on unlabeled data and the softmax outputs of the auxiliary network, and takes the average as the unsupervised loss.$$\begin{aligned} \begin{aligned} O_{\text {total}}&= O_{\text {baseline}} + O_{\text {UA}} + O_{\text {CutMix}} \\&= O\bigl (3BF + T(1 - r)BF + BWH\bigr ) \\&\approx O\Bigl (\bigl (3 + T(1 - r)\bigr )BF\Bigr ) \end{aligned} \end{aligned}$$(10)In Eq. (10), \(O_{\text {baseline}}\) denotes the complexity of a standard training step (one forward \(O(B \times F)\) and backward pass \(O(2B \times F)\)) which costs \(O(3B \times F)\), where \(B\) is the batch size and \(F\) is the number of floating-point operations (FLOPs) per example. Our method adds two additional terms:1.Uncertainty-Aware Module: For the unlabeled portion \((1-r)B\) of the batch, we perform \(T\) stochastic forward passes through the teacher network to estimate uncertainty. It incurs \(O(T(1-r)BF)\), where \(r\) is the labeled data ratio, and r is 0.5.2.Polluted CutMix Preprocessing: Identical in principle to the original CutMix, this step generates a binary mask \(M\) of size \(W \times H\) and performs one pixel-wise weighted addition per image, incurring \(O(B \times W \times H)\) operations. Compared to the baseline’s \(O(B \times F)\), this mask generation and blending overhead is negligible.By choosing a modest number of Monte Carlo samples, we balance uncertainty estimation quality with training efficiency. Because the Polluted CutMix preprocessing cost is minimal, the overall computational complexity of our framework remains on the same order as conventional supervised training. It can be seamlessly integrated with any backbone network.Assumptions and limitations: Our proposed Uncertainty-Aware Polluted CutMix framework is built upon several practical assumptions. First, it assumes that the Polluted CutMix operation preserves diagnostic cues when mixing normal and diseased X-rays, ensuring that the generated samples remain clinically meaningful. Second, the uncertainty estimation derived from the teacher network via Monte Carlo dropout is assumed to reliably reflect the confidence of predictions, enabling effective filtering of ambiguous samples. These assumptions allow the framework to balance the use of labeled and unlabeled data in a stable manner. However, if the mixing process obscures key lesion features or the uncertainty estimates become unstable under strong distribution shifts, performance may degrade. Moreover, the additional stochastic forward passes required for uncertainty estimation increase computational cost compared to conventional training. Future work will address these issues by refining uncertainty modeling and exploring more efficient sampling strategies to further enhance scalability and generalization.Table 1 Performance metrics for different classification tasks.Full size tableNeonatal pulmonary ailment datasetThe Neonatal Pulmonary Ailment (NPA) dataset comprises 2397 clinical cases retrospectively collected from the Women and Children’s Medical Center of Hainan Province between June 2016 and April 2023. This multicohort dataset includes three subgroups: 683 neonates with Neonatal Respiratory Distress Syndrome (NRDS), 886 cases of Aspiration Syndrome (AS), and 828 healthy controls. All bedside chest radiographs were acquired using a standardized Shimadzu DR system, with cases confirmed through rigorous clinical-imaging diagnostic screening. For each case, two complementary data modalities are provided: (1) imaging data, i.e., initial bedside chest X-rays obtained within 0–28 days after birth, and (2) clinical variables, including demographic and perinatal parameters such as gender, gestational age, birth weight, delivery mode (vaginal/cesarean), amniotic fluid status (clear/meconium-stained), and history of asphyxia. Disease annotations follow established clinical-radiographic criteria: NRDS is graded into four stages—Grade I (diffuse fine granular opacities with lower-lung predominance), Grade II (reticulogranular patterns and air bronchograms), Grade III (coarse opacities obscuring cardiac/diaphragmatic borders), and Grade IV (“white lung” appearance)—while AS severity is classified as Mild (bronchovascular thickening), Moderate (patchy consolidations with hyperinflation), or Severe (dense opacities with complications such as pneumothorax). This dataset is designed to address two key clinical challenges. First, NRDS progresses rapidly and is a significant cause of neonatal mortality. By combining temporal X-ray features with perinatal risk factors, the dataset helps improve the early recognition and prediction of NRDS. Second, NRDS is often misdiagnosed as neonatal pneumonia in primary care settings. To reduce this confusion, the dataset explicitly includes distinctive AS subtypes, which help distinguish between these diseases. As the first open-source dataset to pair clinical variables with neonatal pulmonary X-rays, NPA provides a large amount of authentic clinical data while ensuring patient privacy and supporting the development of standardized AI-assisted diagnostic tools.This retrospective study involving human participants was approved by the Medical Ethics Committee of Hainan Women and Children’s Medical Center (approval No. HNWCMC-2024-114). The requirement for informed consent was waived by the Ethics Committee because all patient data were fully de-identified and contained no personally identifiable information. All patient data were fully de-identified prior to analysis to ensure privacy protection, and only anonymized data are shared in the open-access dataset. All research was performed in accordance with relevant guidelines and regulations, and in compliance with the Declaration of Helsinki.Algorithm 1Full size imagePolluted CutMix framework training and evaluation.ExperimentsExperimental detailsImplementation environment and reproducibility: All experiments are conducted in PyTorch 2.1 with CUDA 12.1 on a NVIDIA RTX 4090 GPU. Our method and all baselines share the same codebase, including network backbones, training pipelines, and evaluation routines. We set a global random seed for full reproducibility, perform a comprehensive hyperparameter search for each model, and consistently report the best-tuned results.Backbone configuration: We adopt a simplified DenseNet comprising three dense blocks, each with 16 convolutional layers. Within each block, the growth rate is set to 12 output channels. An initial convolution produces 24 channels, and each bottleneck layer has a size of 4.Training settings: In our experiments, the dataset was divided into training, validation, and test subsets with a ratio of 7:1:2, respectively. All reported evaluation results were obtained on the independent test set, which was strictly excluded from both training and hyperparameter tuning. To prevent data leakage, the splitting was performed on a patient-wise basis, ensuring that images from the same patient did not appear in multiple subsets. This design guarantees a fair and reliable evaluation of the model’s generalization capability. In our work, the batch size for the aspiration syndrome in newborns is set to 8, and for the neonatal respiratory distress syndrome, it is set to 4. In the introduced unsupervised loss, data contaminated with CutMix is treated as unlabeled data, and the labeled data ratio is 0.5. Additionally, we set the EMA decay \(\alpha\) to 0.99. We use the SGD optimizer to update network parameters (weight decay = 0.0001, momentum = 0.9), with an initial learning rate set to 0.01. The learning rate decay strategy is based on the number of iterations, where the learning rate gradually decreases as the number of iterations increases. We compute the decay factor by subtracting the ratio of the current iteration count to the total iteration count from 1, and raising this decay factor to the power of 0.9 to obtain the final learning rate decay coefficient.Cross-validation and dataset split strategy: To ensure fair and robust evaluation, we employed a 5-fold cross-validation scheme across all compared models. The dataset was partitioned into five non-overlapping subsets, maintaining consistent folds for every model to eliminate data-split bias. In each iteration, one fold served as the test set while the remaining four were used for training and validation. The final reported metrics represent the mean performance aggregated across all folds. This strategy provides more reliable generalization estimates and mitigates overfitting risks associated with fixed dataset splits.Uncertainty estimation and thresholding: We use Monte Carlo Dropout with \(T=8\) stochastic forward passes to estimate predictive entropy. The uncertainty threshold H is defined as a monotonically decreasing sigmoid that falls slowly early on, drops sharply near the midpoint, and then levels off—its upper and lower bounds set to the task-specific \(U_{\max }\) and \(U_{\min }\), respectively, estimated via Monte Carlo. This dynamic schedule ensures high tolerance in early training, rapid tightening in mid-training, and stable selection of only the most reliable predictions in later stages.Model Architecture: The proposed system consists of six sequential modules, each responsible for one stage of the data pipeline (see Algorithm 1). (1) Raw Image Preprocessing: Resize and normalize all neonatal chest X-ray images. (2) Polluted CutMix Sample Generation: For each preprocessed image, crop a lesion patch, paste it onto a healthy image, apply basic augmentations, and compute the mixed soft label. (3) Uncertainty Score Computation: Quantify each mixed sample’s noise level via Monte Carlo perturbations and entropy. (4) Unsupervised Loss & Weighting: Compute per-sample consistency loss, map uncertainty to a weight, and perform a weighted average to reduce the impact of noisy samples. (5) Model Training: Train the simplified DenseNet backbone using a combination of supervised loss and uncertainty-aware consistency loss. (6) Inference & Evaluation: Perform inference on original (or lightly augmented) images and compute performance metrics. Note that we also treat the original labeled images as a special case of Polluted CutMix with mixing ratio = 0, and apply the same supervised loss to them.Evaluation metricsWe primarily use four evaluation metrics: accuracy, precision, sensitivity, and specificity. We use these metrics to validate the classification performance of all methods. The formal definitions of these metrics are as follows: \(Accuracy = \frac{TP + TN}{TP + TN + FP + FN}\), \(Precision = \frac{TP}{TP + FP}\), \(Sensitivity = \frac{TP}{TP + FN}\), \(Specificity = \frac{TN}{TN + FP}\). Where TP, FP, TN, and FN denote true positives, false positives, true negatives, and false negatives, respectively. In our experiments, normal cases are considered positive samples. All four metrics were computed on two separate test sets: 248 samples for aspiration syndrome and 245 samples for neonatal respiratory distress syndrome (Table 2).Table 2 Mathematical model comparison.Full size tableComparison to other methods on NPA datasetFirst, we evaluated all methods on the same NPA neonatal pulmonary dataset (see Sect. “Neonatal pulmonary ailment dataset” for details) and compared four standard classification metrics—accuracy, precision, sensitivity, and specificity—defined in Sect. “Evaluation metrics” and summarized in Table 1. To ensure a fair and comprehensive benchmarking, we compared our method not only with CNN-based models (DenseNet, ResNet, VGG, Xception) but also with recent high-performing transformer architectures, including EfficientNet, ConvNeXt, Swin Transformer, and InternImage. We evaluated all methods on the same NPA neonatal pulmonary dataset (see Sect. “Neonatal pulmonary ailment dataset” for details) and compared four standard classification metrics—Accuracy, Precision, Sensitivity, and Specificity—(Sect. “Evaluation metrics”) summarized in Table 1. To quantify experimental variability, we first performed multiple independent training runs (different random seeds, weight initializations, and augmentation orders) and reported mean±standard deviation for each metric. For example, our method achieves an Accuracy of \(80.72\%\pm 0.66\%\) on the AS task, indicating low variability across independent runs. To ensure that the reported improvements are statistically sound rather than random fluctuations, we conducted a comprehensive statistical validation. All models were evaluated using 5-fold cross-validation with identical data splits to guarantee fair and consistent comparisons. For each method, we report the mean and standard deviation of Accuracy across folds, along with the corresponding 95% confidence intervals (CIs) computed using the Student’s t-distribution. To further assess the reliability of the improvements, we conducted paired two-sided t-tests between our method and each baseline using fold-wise Accuracy values. The resulting p values, provided in Table 3, show that almost all comparisons achieved strong statistical significance (p < 0.001), demonstrating that the superiority of our method is not due to random chance. Overall, the narrow 95% CIs and consistently low p values confirm that our method provides statistically significant, stable, and reproducible improvements in Accuracy, reinforcing the robustness of the proposed approach over both CNN- and Transformer-based baselines.Table 3 Cross-validated performance metrics and statistical significance.Full size tableSince the Aspiration Syndrome (AS) task is a binary classification problem, we additionally report the Area Under the Receiver Operating Characteristic curve (AUROC) to provide a threshold-independent evaluation of discriminative ability, which is particularly informative under class imbalance. The AUROC results for all competing methods are summarized in Table 4. Our model achieves an AUROC of 0.8441, demonstrating strong separability between normal and diseased cases and further validating its robustness and generalization ability in the binary classification setting.Table 4 AUROC values for aspiration syndrome (AS) classification.Full size tableBelow, we clarify the fundamental differences among these architectures. Table 2 provides a concise summary of each model’s core modeling idea, Unlike these baselines, which focus solely on network architecture, our approach introduces two mathematical innovations: (1) Polluted CutMix (Eqs. 1–5), which strategically synthesizes minority-class examples by pasting lesion patches onto healthy images; and (2) an uncertainty-aware consistency loss (Eqs. 6–8), which uses Monte Carlo-estimated predictive entropy to filter noisy pseudo-labels.Table 5 Comparison of parameters and computational complexity of different models.Full size tableComplexity analysisIn Table 5, we compare the number of parameters and computational complexity of several backbone networks for a single-channel input image of size 256 \(\times\) 256, measured under identical settings for all models. Note that our DenseNet-based variants (“DenseNet (baseline)” and “Ours”) require only \(\hat{0}\).77 M parameters-an order of magnitude smaller than VGG16 (134.28 M) or ConvNeXt-Tiny (27.82 M)—yet incur a relatively high 19.51 GMac of FLOPs. This drastic reduction in parameter count is achieved by simplifying the DenseNet architecture to three dense blocks with a growth rate of 12, an initial convolution of 24 channels, and bottleneck layers of size 4. All complexity measurements, including those for the baseline DenseNet and competing backbones, were collected using the same batch settings (batch size = 8 for AS and batch size = 4 for NRDS). Despite its extremely compact parameter footprint, this streamlined DenseNet still delivers sufficient representational capacity and compute budget to support our Polluted CutMix augmentation and uncertainty-aware weighting modules.To complement our inference-time analysis (where “Ours” and the DenseNet baseline share the same 0.77 M parameters and 19.51 GMac per image), we profiled training overhead on identical hardware. For the NRDS task (batch = 4), the baseline DenseNet requires 0.117 s per iteration and 8.43 GB of GPU memory. In contrast, our Polluted CutMix + MC Uncertainty pipeline takes 0.213 s and 8.99 GB—a compute time increase of 82% and memory increase of 6.6%. For the AS task (batch = 8), the baseline runs at 0.137 s/iter and 16.71 GB, while our method runs at 0.245 s/iter and 17.28 GB—an increase of 79% in time and 3.4% in memory. These results demonstrate that, although our training strategy increased the time per iteration by about 80%, it adds less than 7% to peak memory usage, a modest trade-off given the substantial performance improvements.Fig. 2Full size imageUncertainty trend changes of aspiration syndrome and neonatal respiratory distress syndrome during training.Fig. 3Full size imageThe accuracy of 100 samples of aspiration syndrome and neonatal respiratory distress syndrome during training.Visualization of uncertainty and accuracy trendsIn our study, to validate the effectiveness of our model, we designed two experiments that illustrate changes in model performance during the training process. Specifically, we trained the model for 41 epochs on the Inhalation Syndrome dataset and 15 epochs on the Neonatal Respiratory Distress Syndrome dataset, providing a more intuitive demonstration of the model’s performance changes throughout the training.The first experiment presents the trend of prediction uncertainty of the model on the aspiration syndrome dataset (Fig. 2b) and the Neonatal Respiratory Distress Syndrome dataset (Fig. 2a) during training. We constructed eight images using polluted CutMix and input these images into the model. These figures illustrate how the model’s prediction confidence for these two conditions evolves with increasing training iterations. Specifically, these plots show the changes in prediction uncertainty for these conditions at each training epoch. From these figures, we observe that as training progresses, the model’s prediction confidence gradually improves, indicating that the model’s discriminative capability is continually enhancing. The decreasing trend in uncertainty with increased training iterations demonstrates that the model is progressively learning and improving its ability to recognize these conditions.Table 6 Performance metrics for different classification tasks.Full size tableThe second experiment illustrates the comparison between the model’s unsupervised labeling results and the manual annotations for 100 randomly selected samples at each training epoch. Figure 3a and b show the labeling accuracy for Neonatal Respiratory Distress Syndrome and Inhalation Syndrome samples, respectively. These figures provide an evaluation of the model’s unsupervised labeling accuracy at each training stage. By examining these plots, we can track the changes in the accuracy of unsupervised labeling across different training phases. The improvement in labeling accuracy as training advances indicates that the model is continually enhancing the quality of unsupervised labeling.In summary, these two experiments showcase the trends in uncertainty and the accuracy of unsupervised labeling throughout the model training process. These visualizations not only provide an intuitive assessment of model performance but also demonstrate the model’s progress and potential in practical applications.Explainable AI and prediction analysisFig. 4Full size imageGrad-CAM visualizations for neonatal pulmonary cases. (a-d) show example model attention maps for Aspiration Syndrome (AS) and Neonatal Respiratory Distress Syndrome (NRDS). “Incorrect” indicates the model’s prediction did not match the ground-truth label.To enhance the interpretability of our proposed model, Grad-CAM was employed to visualize discriminative pulmonary regions associated with predictions. As illustrated in (Fig. 4a and c), the model successfully focuses on bilateral lung fields where pathological manifestations of aspiration syndrome (AS) and neonatal respiratory distress syndrome (NRDS) typically occur. These attention distributions demonstrate that the network has learned clinically relevant visual cues rather than spurious correlations, thereby improving model transparency and trustworthiness.To further investigate model behavior, representative examples of both correct and misclassified cases were analyzed, as shown in (Fig. 4b and d). In the AS misclassification (Fig. 4b), the attention completely drifts away from the lung parenchyma and instead concentrates on non-pathological structures such as the spine and ribs, indicating confusion between anatomical context and lesion-related features. In the NRDS misclassification (Fig. 4d), the model fails to capture the diffuse pulmonary opacities and instead focuses on peripheral regions and body contours, which ultimately leads to an incorrect prediction. These failure cases reveal that errors often arise from attention misalignment or limited representation of subtle radiographic cues. By combining correct and incorrect visualization analyses, our approach provides a more comprehensive understanding of both model reliability and its potential limitations in clinical deployment.Ablation studiesFig. 5Full size imageMetric-wise visualization of ablation study results. Comparative performance of DenseNet, Uncertainty-Aware (UA), Polluted CutMix, and the proposed full model across four evaluation metrics (Accuracy, Precision, Sensitivity, and Specificity) on the Aspiration Syndrome (AS) and Neonatal Respiratory Distress Syndrome (NRDS) classification tasks.We use DenseNet as the backbone network and conduct ablation experiments on the current dataset to validate the effectiveness of our method. To do so, we separate the two modules and conduct ablation experiments on the current dataset. We test the effects of using the polluted CutMix and the uncertainty-aware (UA) framework separately on the DenseNet backbone network, and the results are displayed in Table 6 and visualized in Fig. 5. As shown in Table 6, the final results demonstrate that each of our frameworks is effective, achieving significant improvements over the original DenseNet. Moreover, the combined uncertainty-aware polluted CutMix framework significantly improves the baseline. In the following three chapters, we will continue to conduct further ablation experiments on each of the two modules separately.Table 7 Performance metrics for different classification tasks.Full size tableAblation study on polluted-CutMixWe first validate the effectiveness of polluted CutMix. To demonstrate the effectiveness of our polluted CutMix and illustrate the limitations of conventional CutMix, we conduct further ablation experiments by applying our polluted CutMix and regular CutMix to a fully supervised classification network with DenseNet as the backbone. The specific experimental results are shown in Table 7.From Table 7, we observe that polluted CutMix has only a slight improvement over regular CutMix in the case of aspiration syndrome, but it shows a significant enhancement in pulmonary transparency disease. We attribute this to the fact that aspiration syndrome is a binary classification problem, so even if both selected samples are disease cases, their simultaneous processing has minimal impact because they belong to the same category. However, in the case of pulmonary transparency disease, which involves five classifications, the effect is substantial. With four different disease categories (grades I to IV), randomly selecting samples from different disease categories can easily overlap the shaded areas when processed. This overlapping greatly reduces the model’s discriminative accuracy. Therefore, our polluted CutMix addresses this issue by processing normal cases and disease cases one-to-one, resulting in a significant improvement in data augmentation. This ablation experiment demonstrates the effectiveness of our polluted CutMix.Ablation study on the uncertainty-aware frameworkTable 8 Performance metrics for different classification tasks.Full size tableTo mitigate the potential performance degradation caused by extreme cases where the selected area by the bounding box is non-shadowed but labeled as a diseased case, we introduce the Uncertainty-Aware (UA) framework, which is built upon an unsupervised loss. This framework estimates uncertainty through Monte Carlo sampling, executing T random forward passes on the teacher model with random dropout and Gaussian noise applied to each input tensor. By utilizing the mean or weighted average of these samples to estimate the expected value or integral value of the objective function, high-uncertainty data is effectively filtered out. While this effectively excludes low-probability events from the selected bounding boxes, it introduces unsupervised loss. Some might attribute the performance improvement solely to the unsupervised loss. To address this concern, we further conduct an ablation experiment on the uncertainty-aware model, keeping the unsupervised loss unchanged. We retain the traditional mean teacher model19 but set the consistency loss to zero when the iteration count is less than 1000. This is done to accelerate the model’s convergence, especially in the initial training stages when the model may not have yet learned effective features or patterns. Introducing consistency loss at this stage could make the model difficult to train or slow down convergence. The specific experimental results are shown in Table 8.From the table, we observe that while the introduction of unsupervised loss does lead to some improvement, it still lags behind our uncertainty-aware framework. This demonstrates the effectiveness of our uncertainty-aware model. Here, “Only-usp” represents the introduction of unsupervised loss, while “UA” represents the uncertainty-aware framework.Ablation study on addressing class imbalanceTo validate the effectiveness of our method in addressing data imbalance, we compare it with other related methods. The methods we select include conventional imbalance handling methods and three algorithm-level methods:(i)Inverse and Normalization (IN), which weights the loss inversely proportional to class frequency57.(ii)Focal Loss, which adjusts the weighting by giving less weight to easily classified samples58.(iii)Class-Balanced Loss (CB), which reweights the loss inversely proportional to the effective number of classes59.These methods effectively mitigate the bias caused by class imbalance, enabling the model to achieve better performance on imbalanced data. Additionally, we also use a method for handling class imbalance for comparison: Action++60, an improved contrastive learning framework with adaptive anatomical contrast for semi-supervised medical segmentation. Table 9 presents the results on our dataset.Table 9 Performance metrics for different classification tasks.Full size tableTo further evaluate the robustness and data-efficiency of our model under low-sample and imbalanced conditions, we intentionally increased the imbalance in the experimental dataset to simulate data-scarce clinical scenarios. Specifically, for aspiration syndrome, the training dataset initially contains 475 normal cases and 392 diseased cases. We remove 100 diseased cases, resulting in only 292 cases remaining. Similarly, for pulmonary membrane disease, our dataset initially includes 465 normal cases, 125 cases of level 1, 141 cases of level 2, 109 cases of level 3, and 15 cases of level 4. To exacerbate the imbalance in the dataset, we remove 25 level one cases, 30 level two cases, 20 level three cases, and 5 level four cases. We then compare this reduced-sample setup with commonly used data augmentation and CutMix-based strategies, and the results are shown in Table 10.It can be seen that even under this challenging low-sample condition, our method exhibits the smallest performance degradation compared to other approaches, demonstrating strong resilience and adaptability in data-scarce neonatal imaging scenarios.Table 10 Performance metrics for different classification tasks.Full size tableConclusionsIn this article, we introduced a new dataset on neonatal lung disease and proposed a novel Uncertainty-Aware Polluted CutMix framework to address the challenges inherent in medical image imbalance and uncertainty. This framework leverages Polluted CutMix augmentation to alleviate the long-tail distribution and enhance the diversity of training data, while the uncertainty-aware mechanism mitigates the adverse effects of ambiguous or low-confidence regions. Extensive experiments demonstrate that our method surpasses existing fully supervised models, achieving consistent improvements in diagnostic performance.Strengths and contributions: The originality of this work lies in the integration of targeted data augmentation (Polluted CutMix) and uncertainty modeling into a unified framework for neonatal pulmonary image analysis. The proposed framework integrates data augmentation and uncertainty modeling in a unified design, enabling effective learning from limited and imbalanced neonatal X-ray data. It provides a practical pathway to improve diagnostic accuracy and robustness in neonatal pulmonary disease detection.Limitations: Despite its strong performance, the current study is constrained by the dataset’s size and single-center origin, which may limit its generalizability. Potential small-sample bias, hardware dependency, and dataset-induced noise may also affect performance consistency across different imaging setups. Future work will focus on expanding data collection across multiple institutions and exploring domain adaptation techniques to improve robustness under cross-domain settings.Implications and future directions: The proposed approach holds strong potential for integration into computer-aided diagnosis systems for neonatal lung screening, potentially reducing clinical workload and enabling faster, more consistent decisions. Further research will pursue clinical validation and seamless compatibility with existing diagnostic assistance tools in neonatal care, as well as real-time deployment under uncertainty-aware decision-making scenarios. Beyond pulmonary imaging, the proposed Uncertainty-Aware Polluted CutMix framework possesses strong cross-domain adaptability. Its targeted augmentation and uncertainty modeling principles can be readily extended to other biomedical modalities that face similar data imbalance and annotation scarcity issues, such as cardiac X-rays, fetal ultrasound screening, and pneumonia detection. By integrating uncertainty-aware learning with controlled augmentation, this framework offers a generalizable solution for enhancing model robustness and diagnostic reliability across diverse biomedical imaging domains.Data availabilityThe dataset and complete code implementation used in this study are publicly available at GitHub: https://github.com/wu-jia-huan/Neonatal-Pulmonary-Ailment-Dataset1.ReferencesWang, R., Xu, Z., Wang, X., Liu, W., Lukasiewicz, T. C2M-DoT: Cross-modal consistent multi-view medical report generation with domain transfer network. Information Fusion 125, 103442 https://doi.org/10.1016/j.inffus.2025.103442 (2026).Article Google Scholar Xu, Z., Xu, W., Wang, R., Chen, J., Qi, C., Lukasiewicz, T. Hybrid reinforced medical report generation with M-Linear attention and repetition penalty. IEEE Transactions on Neural Networks and Learning Systems 36(2), 2206–2220 https://doi.org/10.1109/TNNLS.2023.3343391 (2025).Xu, Z., Tian, B., Liu, S., Wang, X., Yuan, D., Gu, J., Junyang, C., Lukasiewicz, T., Leung, V. C. M. Collaborative Attention Guided Multi-Scale Feature Fusion Network for Medical Image Segmentation. IEEE Transactions on Network Science and Engineering 11(2), 1857–1871 https://doi.org/10.1109/TNSE.2023.3332810 (2024).Singh, D. P. et al. A comprehensive study of enhanced computational approaches for breast cancer classification: Comparative analysis with existing state of the art methods. Arch. Comput. Methods Eng. 1–29 https://doi.org/10.1007/s11831-025-10414-5 (2025).Singh, D. P. et al. A comprehensive study on deep learning models for the detection of diabetic retinopathy using pathological images. Arch. Comput. Methods Eng. 33,1–30 (2025).Singh, D. P., Kour, P., Banerjee, T. & Swain, D. A comprehensive review of various machine learning and deep learning models for anti-cancer drug response prediction: Comparative analysis with existing state of the art methods. Arch. Comput. Methods Eng. 32,1–25 (2025).Banerjee, T. et al. Pyramidal attention-based t network for brain tumor classification: A comprehensive analysis of transfer learning approaches for clinically reliable and reliable ai hybrid approaches. Sci. Rep. 15, 28669 (2025).Article ADS CAS PubMed PubMed Central Google Scholar Pacal, I. & Banerjee, T. Towards accurate and interpretable brain tumor diagnosis: T-fspannet with tri-attribute and pyramidal attention-based feature fusion. Biomed. Signal Process. Control 113, 108852 (2026).Article Google Scholar Xu, Z., Liu, Y., Xu, G., Lukasiewicz, T. Self-Supervised Medical Image Segmentation Using Deep Reinforced Adaptive Masking. IEEE Transactions on Medical Imaging 44(1), 180–193 https://doi.org/10.1109/TMI.2024.3436608 (2025).Zhang, J., Zhang, S., Shen, X., Lukasiewicz, T., Xu, Z. Multi-ConDoS: Multimodal Contrastive Domain Sharing Generative Adversarial Networks for Self-Supervised Medical Image Segmentation. IEEE Transactions on Medical Imaging 43(1), 76–95 https://doi.org/10.1109/TMI.2023.3290356 (2024).Zhang, S., Zhang, J., Tian, B., Lukasiewicz, T., Xu, Z. Multi-modal contrastive mutual learning and pseudo-label re-learning for semi-supervised medical image segmentation. Medical Image Analysis 83, 102656 https://doi.org/10.1016/j.media.2022.102656 (2023).Article PubMed Google Scholar Yun, S. et al. Cutmix: Regularization strategy to train strong classifiers with localizable features. In Proceedings of the IEEE/CVF International Conference on Computer Vision 6023–6032 (IEEE, 2019).DeVries, T. & Taylor, G. W. Improved Regularization of Convolutional Neural Networks with Cutout. arXiv preprint arXiv:1708.04552 (2017).Zhang, H., Cisse, M., Dauphin, Y. N. & Lopez-Paz, D. Mixup: Beyond Empirical Risk Minimization. arXiv preprint arXiv:1710.09412 (2017).Banerjee, T. et al. A novel unified inception-u-net hybrid gravitational optimization model (UIGO) incorporating automated medical image segmentation and feature selection for liver tumor detection. Sci. Rep. 15, 29908 (2025).Article ADS CAS PubMed PubMed Central Google Scholar Narayan, Y. et al. A comparative evaluation of deep learning architectures for prostate cancer segmentation: Introducing trionixnet with n-core multi-attention mechanism. Arch. Comput. Methods Eng. 1–40 (2025).Kendall, A. & Gal, Y. What uncertainties do we need in Bayesian deep learning for computer vision? Adv. Neural Inf. Process. Syst. 30 (2017).Xu, Z., Yang, R., Xu, Z., Zhang, S., Yang, Y., Liu, W., Xu, W., Chen, J., Lukasiewicz, T., Leung, V. C. M. PCA: Semi-Supervised Segmentation With Patch Confidence Adversarial Training. IEEE Transactions on Network Science and Engineering 12(4), 2473–2486 https://doi.org/10.1109/TNSE.2025.3548416 (2025).Tarvainen, A. & Valpola, H. Mean teachers are better role models: Weight-averaged consistency targets improve semi-supervised deep learning results. Adv. Neural Inf. Process. Syst. 30 (2017).Shi, Z. et al. MLC: Multi-level consistency learning for semi-supervised left atrium segmentation. Expert Syst. Appl. 244, 122903 (2024).Article Google Scholar Amin, S. U., Hussain, A., Kim, B. & Seo, S. Deep learning based active learning technique for data annotation and improve the overall performance of classification models. Expert Syst. Appl. 228, 120391 (2023).Article Google Scholar Amin, S. U., Abbas, M. S., Kim, B., Jung, Y. & Seo, S. Enhanced anomaly detection in pandemic surveillance videos: An attention approach with efficientnet-b0 and cbam integration. IEEE Access (2024).Singh, D. P., Banerjee, T., Kour, P., Swain, D. & Narayan, Y. Cicada (UCX): A novel approach for automated breast cancer classification through aggressiveness delineation. Comput. Biol. Chem. 115, 108368 (2025).Article CAS PubMed Google Scholar Banerjee, T. Towards automated and reliable lung cancer detection in histopathological images using dy-fspan: A feature-summarized pyramidal attention network for explainable AI. Comput. Biol. Chem.118, 108500 (2025).Banerjee, T. et al. A novel hybrid deep learning approach combining deep feature attention and statistical validation for enhanced thyroid ultrasound segmentation. Sci. Rep. 15, 27207 (2025).Article ADS CAS PubMed PubMed Central Google Scholar Zhong, Z., Zheng, L., Kang, G., Li, S. & Yang, Y. Random erasing data augmentation. Proc. AAAI Conf. Artif. Intell. 34, 13001–13008 (2020).Google Scholar Xu, Z., Liu, Y., Xu, G., Lukasiewicz, T. Self-Supervised Medical Image Segmentation Using Deep Reinforced Adaptive Masking. IEEE Transactions on Medical Imaging 44(1), 180–193 https://doi.org/10.1109/TMI.2024.3436608 (2025).Choe, J. & Shim, H. Attention-based dropout layer for weakly supervised object localization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition 2219–2228 (2019).Xu, Z., Liu, Y., Yuan, D., Li, B., Liu, W., Lukasiewicz, T. You Need Glimpse Before Segmentation: Stochastic Detector-Actor-Critic for Medical Image Segmentation. IEEE Journal of Biomedical and Health Informatics 1–14 https://doi.org/10.1109/JBHI.2025.3623194Ghiasi, G., Lin, T.-Y. & Le, Q. V. Dropblock: A regularization method for convolutional networks. Adv. Neural Inf. Process. Syst. 31 (2018).Tokozume, Y., Ushiku, Y. & Harada, T. Between-class learning for image classification. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 5486–5494 (2018).Verma, V. et al. Manifold mixup: Better representations by interpolating hidden states. In International Conference on Machine Learning 6438–6447 (PMLR, 2019).Summers, C. & Dinneen, M. J. Improved mixed-example data augmentation. In 2019 IEEE Winter Conference on Applications of Computer Vision (WACV) 1262–1270 (IEEE, 2019).Berahmand, K., Daneshfar, F., Salehi, E. S., Li, Y. & Xu, Y. Autoencoders and their applications in machine learning: A survey. Artif. Intell. Rev. 57, 28 (2024).Article Google Scholar Goodfellow, I. et al. Generative adversarial networks. Commun. ACM 63, 139–144 (2020).Article Google Scholar Hu, H., Wang, X., Zhang, Y., Chen, Q. & Guan, Q. A comprehensive survey on contrastive learning. Neurocomputing.610, 128645 (2024).Li, X. et al. Contrastive learning of graphs under label noise. Neural Netw. 172, 106113 (2024).Article PubMed Google Scholar Xu, Z., Wang, H., Yang, R., Yang, Y., Liu, W., Lukasiewicz, T. Aggregated Mutual Learning between CNN and Transformer for semi-supervised medical image segmentation. Knowledge-Based Systems 311, 113005 https://doi.org/10.1016/j.knosys.2025.113005 (2025).Article Google Scholar Banerjee, T. Electromagnetic interaction algorithm (EIA)-based feature selection with adaptive kernel attention network (akattnet) for autism spectrum disorder classification. Int. J. Dev. Neurosci. 85, e70034 (2025).Article CAS PubMed Google Scholar Banerjee, T., Singh, D. P. & Kour, P. Advances in deep neural, transformer learning, and kernel-based methods for diabetic retinopathy detection: A comprehensive review. Arch. Comput. Methods Eng. 1–49 https://doi.org/10.1007/s11831-025-10376-8 (2025).Banerjee, T. Comparing bipartite convoluted and attention-driven methods for skin cancer detection: A review of explainable ai and transfer learning strategies. Arch. Comput. Methods Eng. 1–25 https://doi.org/10.1007/s11831-025-10379-5 (2025).Khan, A. A., Chaudhari, O. & Chandra, R. A review of ensemble learning and data augmentation models for class imbalanced problems: Combination, implementation and evaluation. Expert Syst. Appl. 244, 122778 (2024).Article Google Scholar Duan, W., Hu, N. & Xue, F. The information content of financial statement fraud risk: An ensemble learning approach. Decis. Support Syst. 182, 114231 (2024).Article Google Scholar Yelleni, S. H. et al. Monte Carlo dropblock for modeling uncertainty in object detection. Pattern Recognit. 146, 110003 (2024).Article Google Scholar Suresh, S., Si, Z., Anderson, S., Kaess, M. & Mukadam, M. Midastouch: Monte-carlo inference over distributions across sliding touch. In Conference on Robot Learning 319–331 (PMLR, 2023).Amin, S. U., Taj, S., Hussain, A. & Seo, S. An automated chest x-ray analysis for covid-19, tuberculosis, and pneumonia employing ensemble learning approach. Biomed. Signal Process. Control 87, 105408 (2024).Article Google Scholar Ul Amin, S. et al. EADN: An efficient deep learning model for anomaly detection in videos. Mathematics 10, 1555 (2022).Article Google Scholar Huang, G., Liu, Z., Van Der Maaten, L. & Weinberger, K. Q. Densely connected convolutional networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 4700–4708 (2017).Dosovitskiy, A. et al. An Image is Worth 16x16 words: Transformers For Image Recognition at Scale. arXiv preprint arXiv:2010.11929 (2020).He, K., Zhang, X., Ren, S. & Sun, J. Deep residual learning for image recognition. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 770–778 (2016).Chollet, F. Xception: Deep learning with depthwise separable convolutions. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition 1251–1258 (2017).Simonyan, K. & Zisserman, A. Very Deep Convolutional Networks for Large-Scale Image Recognition. arXiv preprint arXiv:1409.1556 (2014).Tan, M. & Le, Q. Efficientnet: Rethinking model scaling for convolutional neural networks. In International Conference on Machine Learning 6105–6114 (PMLR, 2019).Liu, Z. et al. A convnet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition 11976–11986 (2022).Liu, Z. et al. Swin transformer: Hierarchical vision transformer using shifted windows. In Proceedings of the IEEE/CVF International Conference on Computer Vision 10012–10022 (2021).Wang, W. et al. Internimage: Exploring large-scale vision foundation models with deformable convolutions. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition 14408–14419 (2023).Cao, K., Wei, C., Gaidon, A., Arechiga, N. & Ma, T. Learning imbalanced datasets with label-distribution-aware margin loss. Adv. Neural Inf. Process. Syst. 32 (2019).Lin, T.-Y., Goyal, P., Girshick, R., He, K. & Dollár, P. Focal loss for dense object detection. In Proceedings of the IEEE International Conference on Computer Vision 2980–2988 (2017).Cui, Y., Jia, M., Lin, T.-Y., Song, Y. & Belongie, S. Class-balanced loss based on effective number of samples. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition 9268–9277 (2019).You, C. et al. Action++: Improving semi-supervised medical image segmentation with adaptive anatomical contrast. In International Conference on Medical Image Computing and Computer-Assisted Intervention 194–205 (Springer, 2023).Download referencesAcknowledgementsWe used the test dataset external test dataset from the First Affiliated Hospital of Hainan Medical University, and we thank Director Jianqiang Chen for providing the cases.FundingThis work was supported by the National Natural Science Foundation of China under the grant 62276089, by the Natural Science Foundation of Tianjin, China, under the grants 24JCJQJC00200 and 24JCQNJC01230, by the Natural Science Foundation of Hebei Province, China, under the grant F2024202064, by the Science Research Project of Hebei Education Department, China, under the grant BJ2025004, by the Ministry of Human Resources and Social Security, China, under the grant RSTH-2023-135-1, and by the S&T Program of Hebei under the grant 24464401D.Author informationAuthors and AffiliationsState Key Laboratory of Intelligent Power Distribution Equipment and System, School of Artificial Intelligence, Hebei University of Technology, Tianjin, ChinaJiahuan WuHebei Engineering Research Center of Brain-Computer Intelligent Fusion Technology, School of Health Sciences and Biomedical Engineering, Hebei University of Technology, Tianjin, ChinaJiahuan Wu, Zhenghua Xu & Runhe YangState Key Laboratory of Intelligent Power Distribution Equipment and System, School of Health Sciences and Biomedical Engineering, Hebei University of Technology, Tianjin, ChinaZhenghua Xu & Runhe YangDepartment of Radiology, Hainan Women and Children’s Medical Center, Hainan, ChinaBing ZhuThe Third People’s Hospital of Longgang District Shenzhen, Shenzhen, ChinaYuefu ZhanThe Seventh People’s Hospital of Chongqing, No. 1, Village 1, Lijiatuo Labor Union, Banan District, Chongqing, ChinaYuefu ZhanLonggang Institute of Medical Imaging, Shenzhen, ChinaYuefu ZhanAuthorsJiahuan WuView author publicationsSearch author on:PubMed Google ScholarZhenghua XuView author publicationsSearch author on:PubMed Google ScholarRunhe YangView author publicationsSearch author on:PubMed Google ScholarBing ZhuView author publicationsSearch author on:PubMed Google ScholarYuefu ZhanView author publicationsSearch author on:PubMed Google ScholarContributionsAll authors contributed to the study conception and design. Material preparation, data collection and analysis were performed by Yuefu Zhan, Bing Zhu, Runhe Yang. The first draft of the manuscript was written by Jiahuan Wu and all authors commented on previous versions of the manuscript. All authors read and approved the final manuscript.Corresponding authorsCorrespondence to Jiahuan Wu, Zhenghua Xu or Yuefu Zhan.Ethics declarationsCompeting interestsThe authors declare no competing interests.Ethics approval and consent to participateEthical approval was obtained from the [Medical Ethics Committee of Hainan Provincial Women and Children’s Medical Center, HNWCMC-2024-114].Additional informationPublisher’s noteSpringer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations.Supplementary InformationSupplementary Information. (download PDF )Rights and permissionsOpen Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/.Reprints and permissionsAbout this article