Answers are grounded in this paper and cite their sources.
Timely and consistent assessment of myopia-related fundus pathology remains clinically challenging. Color fundus photography (CFP) and optical coherence tomography (OCT) provide modality-specific information on fundus appearance and retinal structure. This study developed and externally evaluated MVFANet, a clinically guided imaging classification framework comprising separate CFP- and OCT-based classification pathways. This prospective single-center cohort included 179 participants. A total of 343 eyes contributed 4,127 source CFP images (12.03 images per eye), and 314 eyes contributed 4,526 OCT images (14.41 images per eye). Nine-subfield retinal thickness measurements derived from the corresponding OCT examinations were used as auxiliary supervision during OCT model training. Internal performance was estimated using five-fold stratified cross-validation at the patient level. All images, examinations, and both eyes from the same participant were retained within the same fold, and each participant contributed predictions only from the fold in which that participant was not used for model training. Within each training fold, adaptive loss reweighting was applied to dynamically adjust class-specific loss contributions and mitigate majority-class dominance; the held-out out-of-fold data were not used to determine the loss weights or for model selection. For internal assessment, image probabilities were averaged within each eye; CFP predictions were assigned by the highest mean class probability, whereas OCT atrophy and traction were classified independently using fixed decision thresholds. Confidence intervals were calculated using patient-level clustered bootstrap resampling. Independent public datasets were used exclusively for external evaluation after eligibility screening and independent regrading according to the study-specific label definitions. Based on the concatenated patient-disjoint out-of-fold predictions, MVFANet achieved an eye-level accuracy of 88.92% (95% CI: 85.17–92.42), an MCC of 0.8536, a Macro F1 score of 0.8908, and Cohen’s κ of 0.8535 for CFP classification. For OCT-based assessment, atrophy detection achieved an accuracy of 81.85% (95% CI: 76.85–86.67), an MCC of 0.6447, a Macro F1 score of 0.8127, and Cohen’s κ of 0.6297, whereas traction detection achieved an accuracy of 88.85% (95% CI: 84.59–92.85), an MCC of 0.7518, a Macro F1 score of 0.8726, and Cohen’s κ of 0.7461. In the external evaluations, CFP classification achieved an accuracy of 84.25% and a Macro F1 score of 0.8415; OCT-based atrophy and traction detection achieved accuracies of 77.71% and 83.36% and Macro F1 scores of 0.7703 and 0.8133, respectively. Grad-CAM-guided occlusion produced larger probability decreases after masking highly activated regions than after masking random or low-activation regions, supporting the relevance of these regions to the model’s decision process. MVFANet demonstrated modality-specific classification performance under the study’s evaluation conditions. Retinal thickness measurements served as auxiliary supervision during OCT training rather than as independent inference-time inputs. Potential clinical decision-support applications require further validation.
No takes yet. Share an insight, caveat, or question.
Peng et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: