PM-Score-Net: A Disjoint-Label Multi-Task Learning Framework for Automated Continuous Severity Assessment of Myopic Maculopathy

This Code Ocean Capsule provides the pre-trained models, inference scripts, and evaluation metrics to reproduce the results of the study: "PM-Score-Net: A Disjoint-Label Multi-Task Learning Framework for Automated Continuous Severity Assessment of Myopic Maculopathy". This study proposes a novel Multi-Task Learning (MTL) framework named PM-Score-Net, designed to handle disjoint datasets (Classification and Segmentation) using a disjoint-label knowledge assembly strategy. The methodology is centered on a Swin Transformer backbone integrated with a Balanced Supervised Contrastive Loss (SCL) to address class imbalance and diagnostic ambiguity. This capsule allows for the verification of the final optimized model and reproduces key results presented in the study, specifically: 1. The generation of the continuous "qPM-Score" for granular digital biomarker assessment. 2. The state-of-the-art classification performance (Macro F1: 0.7615) on the MMAC 2023 cohort. 3. The 6% innovation gain achieved through disjoint-label multi-task learning. 4. The statistical clinical agreement (Spearman: 0.9062, QWK: 0.8964) and generalizability on the external PALM dataset (100% specificity).