The induction of these organoids typically requires over two months of culture and often results in variable quality, making the process both time-consuming and resource-intensive. To address this bottleneck, the researchers trained a convolutional neural network using phase-contrast images taken during the early stages of organoid development. The model achieved 79% accuracy in predicting pituitary cell differentiation at day 40 using images from day 9, demonstrating its potential to guide experimental decisions before committing to lengthy protocols. Unlike earlier studies that relied on concurrent training and evaluation data, this model forecasts long-term differentiation outcomes based on early-stage imaging, offering a rare predictive capability in organoid biology.

To understand how the model made its predictions, the team applied Grad-CAM, a visualization technique that highlights image regions contributing most to the model's decisions. This analysis revealed that the surface morphology of the organoids—specifically features such as budding patterns and surface texture—was a key determinant of success. While successful organoids tended to show small budding areas and slightly rough surfaces, failed ones exhibited smooth or irregularly rough textures, often associated with mislocalized neural or retinal cells. These morphological cues appeared before molecular markers of differentiation, suggesting that visible structural features can serve as early indicators of developmental potential.

The model's performance was compared with predictions made by experienced researchers, and it consistently outperformed human assessments, particularly at earlier stages. This advantage was most pronounced on day 9, when human predictions were less reliable. The model was also validated across multiple iPS cell lines, confirming its robustness and generalizability beyond the original training data.

To further investigate the biological basis of organoid quality, the researchers conducted RNA sequencing and immunofluorescence analyses. While gene expression profiles of "fail" and "success" organoids were largely similar at early stages, differences in cell type composition and spatial organization became more pronounced over time. Notably, successful organoids more frequently developed oral ectoderm layers on their surface--an essential feature for pituitary differentiation—while failed organoids often showed an overrepresentation of unrelated neural or retinal cells.

The team also examined technical factors influencing model performance. They found that higher-resolution images and larger training datasets improved prediction accuracy, while deviations in focal position during imaging significantly reduced reliability. These findings highlight the importance of standardized imaging protocols for effective deployment of machine learning in organoid research.

By enabling early, non-invasive assessment of organoid potential, this model offers a practical solution for improving the efficiency and reproducibility of organoid-based studies. In a field where long culture periods and inconsistent outcomes have traditionally limited scalability, this approach represents a significant step toward automated, high-throughput organoid production. The machine learning platform is expected to be applicable to other organoid systems and may contribute to advances in regenerative medicine, disease modeling, and drug discovery.

• doi: 10.1016/j.crmeth.2025.101119