Physics-informed deep learning enables reliable and scalable organoid quantification for drug screening via OCT

Patient-derived organoids (PDOs) hold transformative potential for personalized medicine by recapitulating patient-specific drug responses. While Optical Coherence Tomography (OCT) is ideal for monitoring these responses, its translation into high-throughput screening (HTS) is hindered by a segmentation accuracy-throughput bottleneck. Existing solutions fail to meet clinical demands: accurate 3D approaches are prohibitively slow, while emerging foundation models lack sensitivity to minute, low-contrast OCT targets. Conversely, fast 2D models suffer from background noise and unstable performance across varying scales. To bridge this gap, we propose DICE-2DSeg, a physics-informed, graph-enhanced framework. By synergizing OCT-inspired intra-slice coherent enhancement with graph-based inter-slice context aggregation, our method ensures robust quantification. Validated on 93 volumes across diverse cancer types and drugs, DICE-2DSeg demonstrates exceptional robustness. Specifically, our high-throughput variant achieves a 14-fold speedup over nnUNet3D while retaining 93.65% of its accuracy. Crucially, it exhibits superior multi-scale consistency, establishing a new state-of-the-art for challenging drug-responsive remnants (0-100 μm) while maintaining high fidelity for massive clusters (> 100 μm). By resolving the conflict between precision, scalability, and scale-invariance, DICE-2DSeg provides a technical enabling step for automated, large-scale PDO drug screening.