IntroductionThe computer vision (CV) model has shown significant potential in clinical applications by enabling detailed analysis of complex visual information from cell images1. Recently, vision transformer (ViT) has made breakthrough progress in the field of CV, signaling a transition from the Convolutional Neural Network to the Transformer backbone2,3. ViT utilizes a self-attention mechanism to capture long-distance relationships in an image, enabling it to understand global dependencies in the data4. This ability to capture the overall structure of biomedical images is why we selected ViT for this work. CV models using two-dimensional (2D) biomedical images have delivered impressive predictive capabilities in tasks such as detecting cell death5, segmenting cell nuclei6, and localizing subcellular protein7—achievements that are difficult with manual analysis. However, the emergence of physiologically relevant three-dimensional (3D) models like spheroids8,9 and organoids10, underscores the urgent need for the development of advanced 3D imaging techniques and novel cell morphology analysis algorithms in CV. Currently, while drug screening based on phenotypes or statuses from cell images has gradually been applied, there has been less focus on identifying drug-induced liver injury (DILI). This may be due to the substantial metabolic differences between humans and animals11, making it difficult to reflect the actual effects of compounds. For example, traditional preclinical safety trials of 150 drugs reported predictive accuracies of only 63% and 43% in non-rodent and rodent animals, respectively, with the lowest accuracy observed in the hepatobiliary system12. Preclinical identification of DILI-risk compounds remains a challenge in drug discovery13,14, emphasizing the need for alternative in vitro strategies to assess hepatotoxicity compounds and generate reliable data for CV model development. The aim of this work is to develop an expeditious tool based on a CV model for preclinical even clinical drug safety assessment.Organoid culture technology provided new experimentally tractable, physiologically relevant models of human pathologies and subsequent drug screening15. Human liver organoids (HLOs) offer distinct advantages over HepG2 spheroids, as they comprise both hepatic parenchymal and non-parenchymal cells, reflecting accurate intercellular interactions. As 3D multicellular clusters, HLOs carry a cytochrome P450 system involved in drug metabolism, and preserve the phenotype and function of hepatocytes longer than primary human hepatocytes (PHHs)16,17. Recent studies highlight the potential of artificial intelligence (AI)-driven image processing to explore the strong correlation between organoid morphology and compound toxicity or disease status18,19. Therefore, the organoid model is not only a viable alternative to the animal model but also a promising tool primed for assessing DILI risks through morphological analysis. In image analysis, dyes are frequently used to highlight cell features, and CV techniques are then utilized to identify any changes20. Notably, brightfield imaging surpasses fluorescence imaging in several aspects: real-time capabilities, non-destructive nature, and the absence of additional sample processing requirements. Furthermore, brightfield imaging excels in information retrieval due to its high capacity, richness, and depth, all while being cost- and time-effective. To capture the 3D features of the organoid model, we applied 3D video processing principles and developed a CV model based on image-spatial-temporal coding layers to extract spatiotemporal information from high-content screening (HCS). Herein, we developed an evaluation system, named DILITracer, capable of predicting the clinical DILI of compounds based on the HLO technology platform and an AI-assisted algorithm for data analysis. The model achieved an impressive overall accuracy of 82.34%, with particularly high accuracy (90.16%) in identifying non-DILI compounds.To our knowledge, DILITracer is the first model able to categorize hepatotoxicity levels (no, less, or most DILI levels) rather than merely dictating hepatotoxicity. It is simple, non-destructive, and low-cost, with rich information extracted, making it ideal for high-throughput DILI risk evaluation. Our endeavor also represents a significant advancement in compliance with the principles of the 3Rs (Replacement, Reduction, and Refinement). In summary, our innovative AI model utilizes clinical data as an endpoint categorization label, providing a rapid and simple approach to accurately screen compounds with potential clinical liver injury effects.ResultsThe strategy for the DILI-level evaluation system based on the morphology of HLO under brightfieldAs shown in Fig. 1, our approach consists of two stages: system construction and system application: (1) In the stage of system construction: we ensured that the HLOs were in a “drug-ready” state. We also selected 30 structurally and functionally representative compounds with known levels of DILI from DILIrank database21, including four pairs of toxic drugs and their non-toxic structural analogs (troglitazone & pioglitazone, tolcapone & entacapone, nefazodone & buspirone, trovafloxacin & levofloxacin), as well as 22 drugs known to cover known DILI mechanisms22,23, (such as mitochondrial injury, reactive metabolites, biliary transport inhibition, and immune responses) for drug testing. It is noteworthy that the Food and Drug Administration (FDA) DILIrank database categorizes compounds into different degrees of hepatotoxicity based on clinical data, confirming the close alignment of our model with clinical reality during the prediction process. Throughout the testing process, we continuously collected brightfield images of the dosed HLOs using a HCS imager. We then used the DILI levels (No, Less, and Most) of the tested compounds as labels that were added to a series of brightfield images of the corresponding HLOs to generate image sequence DILI-level data pairs. Finally, we trained an AI model to predict the DILI level based on the image sequences to learn the relationship between the image sequences and the DILI severity; (2) In the stage of system application: HLOs, also in the “drug-ready” state, were activated by compounds with unknown DILI severity and then HLO brightfield images were continuously collected during the testing process. The corresponding sequence of brightfield images was input into the AI model to obtain the predicted value of DILI level. The model of this work exhibited an overall accuracy of 82.34%, with a particularly impressive performance in the vNo-DILI-concern category, where it achieved an accuracy of 90.16% (Table 1). This highlights the model’s exceptional ability to identify compounds with no DILI risk, ensuring a high degree of reliability in distinguishing non-hepatotoxic compounds. More detailed comparisons will be discussed in the subsequent sections.Fig. 1: The workflow of the development of the DILI-level prediction model.a The construction of the DILI experimental platform based on human liver organoids for collecting daily brightfield images treated with different compounds under different z-axis. b The establishment of DILI-level prediction model using different brightfield images of liver organoids with spatiotemporal information. c The application of AI model for predicting DILI level of compounds with unknown hepatoxicity.Full size imageTable 1 Predictive performance metrics comparison between two in vivo 3D platformsFull size tableStable establishment of the DILI toxicity testing platform using two distinct 3D liver modelsTo explore the suitability of liver models with varying levels of complexity for DILI toxicity testing platforms, we established a single-type cell 3D model (HepG2 spheroid) and a multi-type cell 3D model (liver organoid). Specifically, HepG2 spheroids and HLOs were each exposed to 30 compounds with or without hepatotoxicity. The levels of ALB from the supernatant and cellular activity (ATP) from the spheres were further assessed at the end of Day 3 to validate the reliability of the system (Figs. 2b, d, f and 3b, d, f). For detailed results regarding changes in ALB and ATP levels of two in vitro 3D models under the treatment of 30 compounds, please refer to Supplementary Figs. 1–4. The brightfield images across different time series and different z-axis orientations were collected daily to generate image data for morphological analysis (Supplementary Figs. 5 and 6). Taking the HLO-based DILI toxicity platform as an example, the significant difference by compounds classified at different levels of liver toxicity potency could be observed. When treated with chlorpheniramine, labeled “No-DILI” by DILIrank, HLOs still increased in diameter and developed into a typical translucent hollow sphere with clear boundaries. In contrast to non-hepatotoxic compounds, Gefitinib-stimulated HLOs, labeled as “Most-DILI”, underwent cell death, failed to maintain their original spherical structure, and disintegrated by the end of Day 3. The state of HLOs treated with Simvastatin (with a label of “Less-DILI”) was between No- and Most-DILI, i.e., HLOs showed growth inhibition but their morphology was still in the form of a complete sphere. Overall, we provided a robust biological basis for the subsequent development of DILI risk prediction models (Fig. 2a, c, e).Fig. 2: DILI toxicity testing platform based on human liver organoid models.a Brightfield morphology changes (12 original images and 1 stacked image) of liver injury in human liver organoids exposed to Chlorpheniramine (with a label of No-DILI levels) from Day 0 to Day 3. b ALB inhibition (P = 0.6262, t = 0.5268) and ATP inhibition (P = 0.8040, t = 0.2652) of liver injury in human liver organoids exposed to Chlorpheniramine at end of Day 3. c Brightfield morphology changes (12 original images and 1 stacked image) of liver injury in human liver organoids liver organoids exposed to Simvastatin (with a label of Less-DILI levels) from Day 0 to Day 3. d ALB inhibition (P = 0.0844, t = 2.285) and ATP inhibition (P = 0.0036, t = 6.110) of liver injury in human liver organoids exposed to Simvastatin at end of Day 3. e Brightfield morphology changes (12 original images and 1 stacked image) of liver injury in human liver organoids exposed to Gefitinib (with a label of Most-DILI levels) from Day 0 to Day 3. f ALB inhibition (P = 0.0191, t = 3.800) and ATP inhibition (P