MyoSpectrumNet: a unified deep learning framework for pathologic myopia spectrum assessment using heterogeneous multi-source fundus datasets

Introduction Pathologic myopia (PM) is a major cause of irreversible visual impairment, yet existing artificial intelligence studies have largely focused on isolated tasks and single-source datasets. Public fundus datasets relevant to PM provide complementary but only partially aligned supervision, creating a need for a unified framework that can support screening, severity assessment, and lesion interpretation across heterogeneous sources. Methods We developed MyoSpectrumNet using HPMI, PALM, and MMAC. The source-conditioned multi-task checkpoint was evaluated on held-out partitions from known sources. Additional task-specific SE and lesion comparators were trained on the official MMAC partitions, with hyperparameters selected only within the training records and single-view inference used for the primary matched-split comparisons. Four-view TTA was retained as a within-MyoSpectrumNet sensitivity analysis. Results Under single-view inference, the task-specific baselines and MyoSpectrumNet yielded HPMI PM AUCs of 0.987 and 0.985, PALM PM AUCs of 0.996 and 0.993, and MM-grading QWK values of 0.907 and 0.907, respectively. The ConvNeXt-T SE baseline achieved an MAE of 1.006 D (95% CI 0.873–1.147) and Pearson r of 0.801, versus a MyoSpectrumNet single-view MAE of 1.259 D. The DeepLabV3-ResNet50 lesion baseline achieved an overall Dice of 0.397 (95% CI 0.324–0.468), versus 0.472 for the MyoSpectrumNet interpretation checkpoint under single-view inference. These auxiliary cross-model comparisons remain descriptive because the original MyoSpectrumNet auxiliary outputs were retained only as aggregate summaries. Conclusion MyoSpectrumNet integrates PM screening, MM grading, auxiliary SE estimation, and lesion-aware visual evidence within one heterogeneous multi-source framework. Task-specific comparisons illustrate the trade-offs of joint learning: it did not improve every output, while the integrated framework retained multiple clinically complementary outputs in one model. The complete multi-output framework has not been externally validated on an unseen dataset, and no universal task-specific superiority claim is made.