Single deep learning model spans pathologic myopia screening, grading and lesion detection

Myopic maculopathy is projected to grow alongside the global rise in myopia, and artificial intelligence (AI) has become one of the main research areas for catching it earlier.² Most AI studies of PM, however, have focused on a single task trained on a single dataset. That leaves clinicians with a patchwork of tools: one algorithm for screening, another for grading and perhaps a third for lesions.

A new study published in Frontiers in Medicine tries to bring these together. Investigators at The First Hospital of Putian in Fujian, China, developed MyoSpectrumNet, a deep learning framework designed to assess the full PM spectrum from a color fundus photograph.¹

Stitching together partial labels

The main technical problem was the data. The 3 public datasets used, HPMI, PALM and MMAC, each carry different and only partly overlapping annotations. HPMI (4,011 images) mainly separates high myopia from PM. PALM (1,200 images) adds PM labels alongside optic disc, fovea and atrophy annotations. MMAC supplies MM grades, SE values and segmentation masks for choroidal neovascularization (CNV), Fuchs spots and lacquer cracks.¹

To combine them, the team used a shared ConvNeXt-T encoder with a source-conditioned adapter that adjusts processing according to the dataset of origin. Separate heads handle PM, MM and SE, and a lesion branch is joined to them through gated fusion. A masked training loss meant each image taught the model only the tasks it had labels for.¹

Matching dedicated models on screening and grading

Each output was compared with a separately trained task-specific baseline on matched data splits, using single-view inference. On the 2 headline clinical tasks, the joint model kept pace. PM screening AUCs were 0.985 on HPMI and 0.993 on PALM, compared with 0.987 and 0.996 for the dedicated classifiers. MM grading produced an identical quadratic weighted kappa of 0.907 for both approaches.¹

The model also transferred reasonably well for PM screening alone. Trained on HPMI and tested on PALM with no fine-tuning, it reached an AUC of 0.993. Transfer in the opposite direction was weaker, at 0.969.¹ The modest compute footprint could matter for real-world deployment: about 31 million parameters and roughly 25 ms per single-view inference on a GPU. That is a recurring theme in point-of-care AI, as a recent offline smartphone screening study reported in Ophthalmology Times showed.⁴

Where joint learning gave ground

The authors do not claim the integrated model wins everywhere, which sets this paper apart from much of the AI literature. SE estimation was weaker. The dedicated regressor achieved a mean absolute error of 1.006 D (95% CI, 0.873–1.147), against 1.259 D for MyoSpectrumNet. The investigators say plainly that joint learning did not improve every output.¹

Lesion segmentation went the other way. MyoSpectrumNet’s interpretation checkpoint posted an overall Dice score of 0.472, against 0.397 (95% CI, 0.324–0.468) for a DeepLabV3-ResNet50 baseline. Even so, the authors call this comparison descriptive. Record-level predictions were not retained, so paired statistics were not possible, and the lesion validation set held only 32 records.¹ They also present the lesion maps as visual evidence to support triage, not as autonomous diagnosis.

Caveats before the clinic

Several limitations stand out for readers weighing clinical relevance. MMAC has no separate test partition, so the MM, SE and lesion figures are validation-set estimates. The most severe maculopathy grades (3 and 4) contributed only 25 records. In that subgroup, grade-4 sensitivity was 1.00 for the baseline but 0.77 for MyoSpectrumNet with test-time augmentation, which is a meaningful gap for the eyes most at risk. Patient-level overlap across the public datasets could not be audited, and comparators were trained with a single seed.¹

The authors also emphasize that the complete 4-output framework has not been tested on an external, unseen dataset. They call for prospective, multicenter and workflow-based validation before any clinical use.¹

Integration over superiority

MyoSpectrumNet’s value is integration, not outright superiority on every task. A single model that screens, grades and highlights lesions fits how PM actually presents: as a continuum of structural change, not as separate yes-or-no findings. That matters as myopia emerges earlier in life. Recent work has shown that more than half of highly myopic paediatric eyes already have peripheral retinal lesions, with prevalence rising alongside axial length.³

For now, the study is best read as a proof of concept for multi-source, multi-task learning in myopia. It is also an unusually honest accounting of what that design gains and gives up. Whether a unified approach can hold up across unseen populations, imaging devices and the most advanced grades of maculopathy is the question the next round of research will need to answer.

References

1. Zhang B, Lin F, Li W, Lin B. MyoSpectrumNet: a unified deep learning framework for pathologic myopia spectrum assessment using heterogeneous multi-source fundus datasets. Front Med (Lausanne). 2026;13:1881289. doi:10.3389/fmed.2026.1881289

2. Hutton D. ASU researchers tap AI to help people see more clearly. Ophthalmology Times. October 7, 2024. Accessed October 30, 2026. https://www.ophthalmologytimes.com/view/asu-researchers-tap-ai-to-help-people-see-more-clearly

3. Bijon J. Detecting peripheral retinal lesions in children with myopia. Ophthalmology Times. June 29, 2026. Accessed October 30, 2026. https://www.ophthalmologytimes.com/view/detecting-peripheral-retinal-lesions-in-children-with-myopia

4. Offline smartphone AI accurately detects DR, glaucoma, and AMD in real-world study. Ophthalmology Times. July 3, 2026. Accessed October 30, 2026. https://www.ophthalmologytimes.com/view/offline-smartphone-ai-accurately-detects-dr-glaucoma-and-amd-in-real-world-study