IntroductionThe Human Protein Atlas (HPA) is a pioneering resource that provides a comprehensive map of protein expression and localization across various human tissues and cell types1,2,3. Established to facilitate a deeper understanding of human biology, the HPA integrates extensive data derived from multiple experimental techniques, including immunohistochemistry, transcriptomics, and mass spectrometry. This atlas offers a wealth of information, detailing the spatial distribution of proteins within different cellular contexts, and serves as a crucial reference for researchers investigating the roles of proteins in both health and disease. By characterizing protein expression patterns, the HPA enhances our understanding of the molecular underpinnings of various physiological processes and pathological conditions4,5. HPA fosters collaborative efforts among researchers worldwide, promoting innovation in fields such as cancer research, neurobiology, and regenerative medicine. The ability to visualize protein expression in situ not only aids in the identification of potential biomarkers for disease but also supports the development of targeted therapies that can be tailored to individual patients based on their unique protein profiles6.Cell-wise classification within the HPA is essential for accurately identifying and characterizing different cell types based on their specific protein expression profiles. This classification is crucial for understanding cellular heterogeneity, as it allows researchers to discern how distinct cell types contribute to overall tissue function and disease processes. Accurate cell classification enhances our ability to investigate cellular interactions, signaling pathways, and the functional roles of proteins, ultimately informing the development of novel therapeutic strategies and personalized medicine approaches7,8.Recent advancements in deep learning methods have revolutionized the landscape of cell-wise classification, enabling more sophisticated analyses of complex biological datasets9. Techniques such as convolutional neural networks (CNNs) and transformer architecture have been successfully applied to extract meaningful features from high-dimensional data, resulting in significant improvements in classification accuracy and robustness10. These methodologies facilitate the extraction of intricate patterns in protein expression and localization, thereby enhancing our understanding of cellular dynamics and their implications in health and disease.Despite the extensive amount of information provided by the Human Protein Atlas (HPA), several specific limitations exist within the available datasets that severely impact machine learning model performance. First, the extreme class imbalance problem where nucleoplasm represents 27.4% of samples while rare localizations like mitotic spindle constitute only 1.3%, creating a 21:1 imbalance ratio that leads to biased model predictions favoring dominant classes. Second, the insufficient representation of morphological diversity within each protein localization class, where subtle but biologically significant variations in cellular phenotypes are underrepresented, limiting model generalization to novel cellular contexts. Third, the missing pixel artifacts that frequently occur during high-resolution fluorescence microscopy imaging, particularly at organellar boundaries and in regions with weak protein expression, which create incomplete training examples that degrade classification performance. Fourth, the poor connectivity preservation between neighboring pixels in cellular structures, where traditional augmentation techniques fail to maintain the biological plausibility of organellar networks and protein distribution patterns essential for accurate classification11,12.The primary aim of this paper is to systematically address these four critical limitations through a novel dual-generator GAN architecture that specifically tackles: (1) class imbalance through targeted minority class augmentation, (2) morphological diversity enhancement via biologically-constrained synthetic image generation, (3) missing pixel imputation using secretion-pattern modeling, and (4) connectivity preservation through homeostatic regularization mechanisms. While our methodology is primarily developed and validated on HPA data, the underlying biological principles of cellular organization, protein localization patterns, and organellar structure are universal across mammalian cell types, making our approach broadly applicable to other fluorescence microscopy datasets including immunofluorescence imaging, live-cell imaging, and multi-channel cellular phenotyping studies beyond the HPA domain.Generative Adversarial Networks (GANs) are a powerful class of machine learning models that consist of two neural networks: a generator and a discriminator13. The generator creates synthetic images, while the discriminator evaluates their authenticity against real images. This adversarial process continues until the generator produces images that are indistinguishable from real ones. GANs can be particularly useful for data augmentation, as they can generate diverse and high-quality images that enhance existing datasets. In the context of the Human Protein Atlas (HPA), GANs can help address limitations such as insufficient representation of various cell types and conditions. By generating realistic protein expression images, GANs can enrich the dataset, improve the training of machine learning models, and ultimately lead to more accurate classifications and insights into protein localization and function14,15.However, the application of GANs is accompanied by several challenges. Common limitations include mode collapse, where the model generates a limited variety of images, and issues related to missing pixels and low integrity between neighboring pixels in generated images. These factors can compromise the quality of the augmented data, potentially undermining the performance of downstream classification tasks. Addressing these limitations is essential for maximizing the benefits of GANs in cellular image augmentation, ensuring that generated images maintain high fidelity and accurately represent the biological variability present in real datasets16,17. The main challenge in using GAN architecture is about developing a loss function to handle the augmented images in missing pixels and connectivity between the pixels in the augmented images18,19.GANs address insufficient representation of cell types through targeted minority class augmentation that generates biologically plausible morphological and phenotypic variations within underrepresented classes, fundamentally expanding the diversity space beyond what is inherently present in the original dataset. Our approach operates on the principle that insufficient representation typically stems from sampling limitations during data collection rather than the biological non-existence of cellular states, particularly for rare localization patterns such as mitotic spindle and aggresome that occur during specific cell cycle phases or stress conditions. The SG-Loss function specifically models the continuous spectrum of secretory states by capturing graded protein expression patterns and spatial reorganization dynamics that characterize transitional cellular conditions, enabling the generation of intermediate morphological states that bridge the gaps between the limited discrete examples present in the training data. This mechanism leverages the biological reality that cellular phenotypes exist along continuous gradients of functional states, allowing the generator to interpolate between observed configurations while maintaining adherence to established biological constraints such as organellar volume ratios, connectivity patterns, and protein localization hierarchies. The PG-Loss function complements this approach by enforcing structural integrity constraints that preserve the fundamental organizational principles of cellular architecture, ensuring that generated variations maintain realistic subcellular compartmentalization, organellar relationships, and morphological characteristics specific to each cell type. The dual-generator architecture implements controlled perturbation of protein localization intensity distributions, spatial clustering patterns, and co-localization relationships within minority classes, systematically exploring the biologically feasible parameter space around the limited training examples to generate synthetic instances that capture the natural phenotypic variability expected in these cell populations. The adaptive fusion mechanism dynamically weights the contributions of secretory modeling and structural preservation based on inferred cellular functional states, enabling the generation of synthetic examples across the full spectrum of physiological conditions including quiescent, proliferative, differentiated, and stress-response states that are systematically underrepresented in static imaging datasets. This methodology differs fundamentally from conventional data augmentation techniques that apply geometric transformations to existing images, as our approach generates novel cellular configurations by modeling the underlying biological processes that govern cellular organization, protein trafficking, and functional state transitions, thereby creating synthetic training examples that expand the representational capacity of minority classes while maintaining biological validity and enabling more robust feature learning for rare cell types.Our approach enhances diversity through three complementary mechanisms: (1) Morphological diversity expansion where the SG-Loss generates novel but biologically plausible cellular morphologies by modeling the continuous spectrum of secretory states observed in glandular cells, creating synthetic examples that fill gaps in the morphological space between existing training samples. (2) Protein expression pattern diversification where the dual-generator architecture produces variations in protein localization intensity, spatial distribution, and co-localization patterns while maintaining biological constraints, effectively expanding the representational capacity of minority classes without introducing artifacts. (3) Phenotypic state diversity where the adaptive fusion mechanism dynamically balances between secretory-active and structurally-stable cellular states, generating synthetic cells across the full spectrum of physiological conditions rather than limiting generation to average or dominant phenotypes present in the original dataset.In this paper, we propose an innovative approach to overcome the limitations associated with existing datasets and the challenges faced in using Generative Adversarial Networks (GANs) for cellular image augmentation. We introduce two biologically inspired loss functions: the “Salivary Gland” loss function for effective missing pixel imputation and the “Pituitary Gland” loss function to ensure integrity among neighboring pixels in generated images. By integrating these loss functions into the GAN framework, we enhance the quality of the synthetic images produced, addressing issues such as mode collapse and pixel integrity. Our methodology not only improves the diversity and realism of the generated images but also enriches the training dataset, ultimately leading to more robust cell-wise classification models. Through comprehensive experiments, we demonstrate that our approach significantly increases classification accuracy and enhances the overall performance of deep learning algorithms, paving the way for more effective and reliable cellular analysis in the context of the Human Protein Atlas.The key contributions of the paper are as follows:1.Introduction of the novel Salivary Gland loss function (SG-Loss), which models the graded secretion patterns of acinar cells to address missing pixel imputation in cellular images2.Development of novel Pituitary Gland loss function (PG-Loss), which implements homeostatic regularization to preserve structural integrity across subcellular compartment boundaries3.Demonstration that these specialized loss functions work synergistically to enhance both image restoration and subsequent classification tasks4.Seamless integration of the dual loss functions within a GAN framework optimized for cellular imaging data5.Introducing a new architecture of GANs with a new two-loss function which can defeat other architectures in terms of quality and diversity in the augmented images6.Development of an adaptive weighting mechanism that balances the influence of each loss component based on image characteristicsUnlike existing biologically inspired algorithms, our Dual-Gland approach offers several distinct advantages:1.While methods like Genetic Algorithms and Particle Swarm Optimization use broad biological concepts (evolution, flocking), our loss functions directly model specific cellular processes relevant to the imaging domain2.The SG-Loss and PG-Loss incorporate actual cellular biology knowledge rather than just mimicking biological optimization patterns3.Our approach specifically addresses the unique challenges of cellular imaging: missing data imputation and maintaining biological plausibility4.Unlike standalone bio-metaheuristics, our loss functions are specifically designed to enhance neural network training5.They complement rather than replace gradient-based optimization, allowing for end-to-end trainingThis paper is organized as follows: Section "The related work" presents a comprehensive review of related work, highlighting existing methodologies in cellular image classification and augmentation techniques, particularly focusing on the applications of Generative Adversarial Networks (GANs). In Section "Materials and methods", we detail the materials and methods employed in our study, including the implementation of the proposed biologically inspired loss functions and the architecture of the GAN framework. Section "Results and discussion" provides a thorough analysis of the results, demonstrating the effectiveness of our approach through quantitative metrics and qualitative assessments of generated images. We discuss the implications of our findings in the context of cellular classification and the potential impact on future research. Finally, Section "Conclusion and future work" concludes the paper by summarizing the key contributions of our work and outlining directions for future research, emphasizing the importance of further enhancing data augmentation techniques to support advancements in biomedical imaging and analysis.The related workThis section of the paper presents the different methods of classification of the cellwise in the human protein atlas and also presents the different GANs architectures in image augmentations. This section also presents the research gaps in the previous and recent work.Classification of cellwise in HPAWei Ouyang et al.20 conducted a significant study focused on the challenges of identifying subcellular protein localizations from microscopy images, a task that is relatively straightforward for trained observers but difficult to automate effectively. To tackle this issue, they organized a competition utilizing the Human Protein Atlas image collection, aimed at fostering the development of deep learning solutions for this complex task. The competition revealed several critical challenges, including the presence of highly imbalanced classes and the requirement for models to predict multiple labels for each image. Over three months, 2172 teams participated, demonstrating a wide variety of approaches despite a general trend toward utilizing popular neural network architectures and training techniques. Participants implemented various strategies, including modifications to neural network structures, innovations in loss functions, data augmentation techniques, and the use of pretrained models to enhance performance. The winning models achieved remarkable results, outperforming previous efforts in multi-label classification of protein localization patterns by approximately 20%. These advanced models not only function as classifiers for annotating new images but also serve as feature extractors for measuring pattern similarity, making them versatile tools for a broad spectrum of biological applications.Tahani Alsubait et al.21 have highlighted the significant advancements achieved by deep learning across various domains, particularly in the medical field. These advancements encompass tasks such as image classification and object detection, with profound implications for single-cell classification. Deep learning algorithms have transformed this area by enabling the classification of cellular components and the precise localization of proteins within cells. The vast diversity of cell types and sizes in the human body complicates traditional analysis methods, thereby revealing a critical research gap that deep learning methodologies are beginning to fill. In their study, the authors utilized the Human Protein Atlas dataset, which consists of 87,224 images of single cells, to evaluate the effectiveness of three innovative deep learning architectures: CSPNet, BoTNet, and ResNet. The results demonstrated impressive accuracy of 95%, 93%, and 91%, respectively, underscoring the potential of these algorithms to enhance the analysis of single cells and contribute to advancements in cellular biology.Lina Al-joudi et al.22 emphasized the critical importance of subcellular localization of human proteins in understanding the structural organization of human cells. They noted that proteins played a vital role in cellular functions, with distinct groups localized to specific regions to execute specialized tasks. Understanding these localized functions was essential for identifying various diseases and developing targeted therapeutic interventions. The researchers recognized the increasing significance of imaging analysis techniques in proteomics research. Despite advancements in deep learning algorithms for analyzing microscopy images, classification models faced considerable challenges in achieving optimal performance, particularly due to the prevalent issue of class imbalance in protein subcellular images. To address this challenge, the authors employed both oversampling and undersampling techniques. They utilized a Convolutional Neural Network (CNN) architecture known as GapNet-PL for the multi-label classification task on the Human Protein Atlas Classification (HPA) dataset. Their findings indicated that the Parametric Rectified Linear Unit (PReLU) activation function outperformed the Scaled Exponential Linear Unit (SeLU) activation function across various classification metrics. Specifically, the GapNet-PL model with the PReLU activation function achieved an area under the receiver operating characteristic curve (AUC) of 0.896, an F1 score of 0.541, and a recall of 0.473, underscoring its effectiveness in addressing multi-label classification challenges in proteomics.Yumen et al.23 focused on single classification models for the task of classifying human protein cell images, aiming to identify specific proteins based on various cell types. However, they noted that traditional classifiers typically identified only one protein at a time, while a single cell often contained multiple proteins that were not entirely independent of one another. In their research, they developed a human protein cell classification model utilizing multi-label learning to address this limitation. They analyzed the logical relationships and distribution characteristics among labels to determine the different proteins present in a set of diverse cells, allowing for multiple outputs in the classification task. Using human protein image data, the authors conducted comparative experiments on pre-trained models, specifically Xception and InceptionResNet V2. They optimized these models through data augmentation, channel settings, and structural adjustments. Their results demonstrated that the optimized InceptionResNet V2 model achieved high performance in the classification task, with a final accuracy of 96.1%. This represented a notable improvement of 2.82% compared to the accuracy achieved before optimization.The aforementioned studies have reported impressive overall classification accuracies ranging from 89% to 96.1% on HPA datasets, these high-level metrics mask several fundamental limitations that compromise the practical applicability and robustness of these classification systems. The characterization of classification accuracy as insufficient stems from four critical performance gaps that existing high-performing models fail to address comprehensively. First, severe performance degradation on minority classes where models achieving 96% overall accuracy demonstrate dramatically reduced effectiveness for rare protein localizations, with class-specific performance metrics revealing F1-scores below 0.3 for classes like mitotic spindle, aggresome, and nucleoli fibrillar center despite the impressive overall accuracy statistics. This disparity indicates that high overall accuracy is primarily driven by dominant classes like nucleoplasm and cytosol, while the classification system fails to provide reliable predictions for biologically important but statistically rare cellular phenotypes that are crucial for understanding specialized cellular functions and disease mechanisms.limited generalization capability across diverse experimental conditions where models trained and evaluated on carefully curated HPA datasets under controlled laboratory conditions fail to maintain their reported high accuracy when applied to images from different research laboratories, varying imaging protocols, alternative fluorescent labeling techniques, or different cell culture conditions commonly encountered in real-world research environments. The high accuracy values reported in controlled studies do not translate to robust performance across the full spectrum of biological and technical variability present in routine cellular imaging applications, limiting the practical deployment of these classification systems in diverse research contexts where imaging conditions cannot be standardized to match training data specifications.Insufficient robustness to data quality variations where the reported high accuracy is achieved primarily when using high-quality, artifact-free images with optimal signal-to-noise ratios, proper illumination, and minimal technical artifacts, but performance deteriorates significantly when dealing with common imaging challenges including missing pixels due to photobleaching, poor connectivity between cellular structures due to inadequate resolution, varying illumination conditions, and imaging artifacts that frequently occur in routine laboratory practice. The classification systems demonstrate brittleness when faced with real-world data quality issues that deviate from the pristine conditions assumed during training and evaluation phases.Inadequate handling of class imbalance complexity where existing approaches rely primarily on basic sampling techniques such as oversampling and undersampling without addressing the fundamental challenge that insufficient representation of rare cell types stems from the inherent difficulty of capturing diverse cellular states rather than simple statistical undersampling. The problem extends beyond numerical balance to encompass morphological diversity within each class, where rare classes lack sufficient examples of the natural phenotypic variability expected in these cell populations, leading to overfitted classification models that cannot generalize to novel morphological presentations of the same protein localization patterns.Research gaps in previous classification modelsExisting studies primarily rely on the Human Protein Atlas dataset without addressing its inherent class imbalance problems beyond basic sampling techniques, failing to generate biologically plausible synthetic examples that expand the morphological diversity within minority classes while maintaining adherence to established cellular organization principles.The classification accuracy demonstrates insufficient robustness for practical deployment, where high overall metrics achieved under controlled experimental conditions do not translate to reliable performance across diverse imaging conditions, alternative experimental protocols, or varying data quality scenarios commonly encountered in real-world research environments.The classification models depend on datasets that inadequately represent the full spectrum of biological and technical variability, lacking examples of intermediate cellular states, transitional morphologies, and diverse phenotypic presentations within each protein localization class that are essential for developing robust and generalizable classification systems.Rare protein localization patterns are severely underrepresented in both quantity and morphological diversity, hindering effective model training for these classes and resulting in classification systems that perform well on dominant classes but fail to provide reliable predictions for biologically important but statistically rare cellular phenotypes that are crucial for understanding specialized cellular functions and disease-related protein mislocalization patterns.Related work on image augmentation architectures for cellular imagingImage augmentation in cellular imaging has gained significant traction through the application of advanced architectures, particularly Generative Adversarial Networks (GANs). These techniques address the challenge of limited training data by generating synthetic images that closely resemble real cellular structures, thereby enhancing the performance of deep learning models in various tasks such as segmentation and tracking.Cellular images of different modalities in machine learning are the collection of cellular images of different microscopy techniques24. These cheap, heterogeneous imaging modalities are widely used in image segmentation tasks in practice. However, state-of-the-art methods for cellular image segmentation fail to generalize between different image modalities. To tackle this problem, a novel generative adversarial network is proposed to construct a portable framework for cellular image modality augmentation. A cycle-consistency loss for modality-convolutional neural network-generalized augmentations is designed to improve training sample size and diversity during training time without any post-processing or fine-tuning25. As a use case, the performance of the proposed method for cellular image segmentation between the bright field and phase contrast microscopy modalities is thoroughly evaluated using three public datasets26. Experimental results demonstrate the effectiveness of the proposed framework in improving the generalization across images captured with different imaging modalities. Cell segmentation methods, especially deep learning approaches relying on training with many labeled image examples, have been extensively studied in recent years. Nevertheless, the collected datasets are mostly limited because labeling cellular images is time-consuming and expertise dependent.GANs have been effectively utilized to create realistic synthetic microscopy images, improving data availability for training models27,28. Popular architectures like StyleGAN and various loss functions (e.g., Wasserstein loss) have been identified as effective in augmenting cell microscopy images27. GANs can generate 2D and pseudo-3D images, retaining biological structure integrity, which is crucial for accurate analysis28. Incorporating real augmentations, such as intentionally defocused images, has been shown to outperform traditional computational methods in improving segmentation accuracy. This approach enhances the robustness of models by providing diverse training scenarios that reflect real-world conditions29.Tools like SynCellFactory30 leverage GANs to produce synthetic cell videos, significantly boosting tracking performance in scenarios with sparse data. This generative approach allows for the simulation of complex cellular behaviors, facilitating better model training. While GANs and real augmentation techniques show promise in enhancing cellular imaging, challenges remain, such as the need for high-quality training data and the computational demands of training these models. Balancing synthetic and real data remains crucial for optimal performance in cellular imaging tasks. Table 1 represents a systematic comparison of augmentation techniques for cellular imaging.Table 1 Comparison of augmentation architectures for cellular imaging.Full size tableCritical Analysis of Classification Performance Limitations: While previous studies have reported high classification accuracies of 89–96% on HPA datasets, these achievements are primarily obtained under specific experimental conditions that do not fully represent the complexity of real-world cellular analysis scenarios. The classification accuracy limitations we address stem from three fundamental issues that existing high-performing models fail to overcome: (1) Performance degradation on rare cell types, where models achieving 96% overall accuracy often perform poorly on minority classes like mitotic spindle (1.3%) and aggresome (1.3%), with class-specific F1-scores dropping below 0.3 for these rare phenotypes despite high overall metrics; (2) Limited generalization across different experimental conditions, where models trained on specific imaging protocols, cell culture conditions, or fluorescent labeling techniques fail to maintain their reported high accuracy when applied to images from different laboratories or experimental setups; and (3) Insufficient robustness to data quality variations, where high accuracy is achieved only when using carefully curated, high-quality images, but performance significantly deteriorates when dealing with real-world scenarios involving missing pixels, imaging artifacts, poor signal-to-noise ratios, or varying illumination conditions commonly encountered in routine laboratory practice. These limitations highlight that while existing methods can achieve impressive performance metrics under controlled conditions, they lack the robustness and generalizability required for practical deployment in diverse research environments where cellular imaging conditions and quality vary significantly.Comprehensive Research Gap Analysis and Targeted Improvements: Our Dual-Gland GAN architecture systematically addresses seven critical research gaps identified across existing augmentation methods through innovative biological modeling approaches that target specific limitations outlined in Table 1. First, regarding severe mode collapse with protein localization patterns, which affects traditional GANs31 and DCGAN32 particularly when generating rare cellular phenotypes, our dual-generator architecture fundamentally prevents mode collapse through biological constraint enforcement where the SG-GAN specializes in secretory pattern modeling while PG-GAN focuses on structural integrity, ensuring complementary functional spaces that cannot collapse to producing repetitive or limited cellular morphologies. The adaptive fusion mechanism dynamically balances these generators based on biological state indicators, preventing either generator from dominating the training process and maintaining diverse output generation across all cell types including extremely rare classes.For comprehensive missing pixel recovery, which is poorly handled by most existing methods including WGAN33, CycleGAN36, and ECP-IGANN38 that either ignore missing data entirely or apply simplistic interpolation, our SG-Loss function implements sophisticated missing pixel imputation based on graded secretion patterns observed in salivary acinar cells. This approach models the hierarchical organization of secretory machinery at multiple scales, enabling intelligent reconstruction of incomplete cellular features through multi-scale contextual analysis that captures fine-grained secretory granule details, intermediate organellar relationships, and coarse-scale cellular polarity patterns. The mathematical formulation incorporates Laplacian operators to ensure smooth concentration gradients characteristic of biological protein distribution, creating realistic imputation that respects cellular biology rather than arbitrary pixel filling.Concerning limited diversity for rare cell phenotypes, which plagues methods like Conditional GAN35 and Pix2Pix GAN34 that struggle to generate meaningful variations within minority classes, our approach addresses this through morphological diversity expansion where the SG-Loss generates novel but biologically plausible cellular morphologies by modeling the continuous spectrum of secretory states observed in glandular cells. This creates synthetic examples that fill gaps in the morphological space between existing training samples, effectively expanding the representational capacity of minority classes through controlled perturbation of protein localization intensity distributions, spatial clustering patterns, and co-localization relationships while maintaining adherence to established biological constraints.Addressing poor preservation of biological structures, which is inconsistently maintained in CollaGAN39, MCI-GAN40, and GSIP-GAN37 that focus primarily on visual realism without biological validation, our PG-Loss function enforces organelle-specific constraints through homeostatic regularization that adaptively weights pixel relationships based on subcellular compartment boundaries. The compartmentalization factors reflect quantitative measurements of cellular organization with specific values for within-organelle interactions (1.0), between adjacent organelles (0.5), and distant cellular regions (0.1), ensuring that generated images maintain realistic organellar volume ratios, connectivity patterns, and morphological characteristics derived from stereological measurements of over 1000 glandular cells.Targeting weak preservation of organelle relationships, which affects DCGAN32 and traditional GANs31 that treat cellular structures as independent visual elements, our approach implements comprehensive biological constraint integration through volume constraints that enforce nuclear (15–20%), mitochondrial (10–15%), and endoplasmic reticulum (8–12%) volume fractions, connectivity constraints that preserve organellar network topology essential for cellular function, and morphological constraints that maintain cell aspect ratios within biologically observed ranges (1.2–2.5) based on morphometric analysis of over 500 cells per type.Addressing inconsistent structural preservation across different cellular contexts, which challenges methods like ECP-IGANN38 and CycleGAN36 that lack biological grounding, our mechanistic translation framework ensures that each mathematical component directly corresponds to measurable biological phenomena. The multi-scale feature extraction mathematically represents the hierarchical organization of secretory machinery where fine-scale features correspond to individual secretory granules with specific diameters (0.5–2 μm), intermediate-scale features capture spatial organization relative to Golgi apparatus and endoplasmic reticulum, and coarse-scale features model overall cellular polarization and apical-basal protein concentration gradients.Resolving training instability and complex implementation issues that affect methods like GSIP-GAN37, CollaGAN39, and MCI-GAN40, our adaptive hyperparameter optimization framework dynamically adjusts critical parameters during training based on convergence metrics and biological plausibility indicators. The system implements proportional-integral control for loss weight balancing, biological constraint monitoring that evaluates structural integrity metrics including organellar boundary definition and protein localization coherence, and statistical significance testing that requires p