Development and validation of deep learning for predicting the growth of ovarian cancer organoids

Abstract Background: Organoids have attracted enormous interest in disease modeling, drug screening, and precision medicine. However, developing robust patient-derived organoids (PDOs) was time-consuming, costly, and had low success rates for certain cancer types, which limited their clinical utility. This study aimed to develop an interpretable deep learning-based model to predict the cultivation outcome of ovarian cancer organoids in advance. Methods: Longitudinal microscopy images of 517 ovarian cancer organoid droplets were divided into training ( n = 325), validation ( n = 88), and test ( n = 104) sets. Subsequently, growth prediction models were developed based on four neural network backbones (ResNet18, VGG11, ConvNeXt v2, and Swin Transformer v2), and specific optimization methods were designed for better prediction. Finally, 179 samples from multiple centers were collected for prospective validation, and the gradient-weighted class activation mapping (Grad-CAM) method was used for interpretability analysis of the deep model to reveal the basis of the model’s decisions. Results: The test set showed that the deep learning models could achieve high-performance prediction at the third stage with area under the curve (AUC) values greater than 0.8 for all four models. The homogeneous transfer learning optimization method improved the AUC from 0.833 to 0.884 ( P = 0.0039). In prospective validation, the optimized model achieved an AUC of 0.832, a Brier score of 0.1919 in the calibration curve, and a greater net benefit in the decision curve. Interpretability analysis revealed that the area where organoids are being formed and have already formed is important for prediction. Conclusions: Our developed models achieved satisfactory results in predicting the growth of ovarian cancer organoids. There is potential for further development of the model toward process automation.