Lesion detection with fine‐grained image categorization for myopic traction maculopathy (MTM) using optical coherence tomography
Xingru Huang, Shucheng He, Jun Wang, Shangchao Yang, Yaqi Wang, Xin Ye
Abstract
BACKGROUND: Myopic traction maculopathy (MTM) are retinal disorder caused by traction force on the macula, which can lead to varying degrees of vision loss in eyes with high myopia. Optical coherence tomography (OCT) is an effective imaging technique for diagnosing, detecting and classifying retinopathy. MTM has been classified into different patterns by OCT, corresponding to different clinical strategies. PURPOSE: We aimed to engineer a deep learning model that can automatically identify MTM in highly myopic (HM) eyes using OCT images. METHODS: A five-class classification model was developed using 2837 OCT images from 958 HM patients. We adopted a ResNet-34 architecture to train the model to identify MTM: no MTM (class 0), extra-foveal maculoschisis (class 1), inner lamellar macular hole (class 2), outer foveoschisis (class 3), and discontinuity or detachment of foveal outer hyperreflective layers (class 4). An independent test set of 604 images from 173 HM patients was used to evaluate the model's performance. Classification performance was assessed according to the area under the curve (AUC), accuracy, sensitivity, specificity. RESULTS: Our model exhibited a high training performance for classification (F1-score of 0.953; AUCs of 0.961 to 0.998). In test set, it achieved sensitivities (91.67%-97.78 %) and specificities (98.33%-99.17%) as good as, or better than, those of experienced clinicians. Heatmaps were generated to provide visual explanations. CONCLUSIONS: We established a deep learning model for MTM classification using OCT images. This model performed equally well or better than retinal specialists and is suitable for large-scale screening and identifying MTM in HM eyes.