Deep Learning Models Can Identify Myopic Maculopathy

Researchers have developed deep learning models and systems that can achieve high sensitivity and specificity in identifying lesions of myopic maculopathy, according to research published in Ophthalmology Retina.

Myopic maculopathy, a cause of irreversible vision loss, is difficult for general ophthalmologists to diagnose, the researchers explain. Early detection is crucial. The investigators sought to develop an automated diagnosis system to help identify patients who need care from specialists.

They collected, graded, and labeled 7020 fundus images from 4432 eyes from Advanced Clinical Center for Myopia of the Tokyo Medical and Dental University data.

Lesions were classified according to META-PM study group based on the most serious lesion present: no myopic retinal lesions (Category 0 or C0), tessellated fundus only (C1), diffuse atrophy (C2), patchy atrophy (C3), and macular atrophy (C4). Pathologic myopia was denoted as at least diffuse atrophy or by “plus” signs. Lacquer cracks (LCs), myopic choroidal neovascularization (CNV), and Fuchs’ spot were classified as plus lesions.

The researchers selected 2 images per eye for the training and validation dataset for a total of 5176 images from 2588 eyes of 1566 patients with pathologic myopia. The grouped fundus images were standardized and augmented to increase image heterogeneity.

Four binary models were trained to categorize diffuse atrophy, patchy atrophy, macular atrophy and CNV. Together, they formed a META-PM categorizing system (CS) through the addition of a specific processing layer. Each algorithm was evaluated based on the area under the curve (AUC) of the receiver operating characteristics (ROC) curve, sensitivity, and specificity. The researchers tested the system’s generalization ability on the open-source PALM dataset of fundus images with pathologic myopia and a Singapore Epidemiology of Eye Disease population-based dataset.

The accuracy of diffuse atrophy model was 89.25%. AUC was 0.970 with sensitivity of 84.44% and specificity of 94.50%. For the patchy atrophy algorithm, the model accuracy was 94.86% with an AUC of 0.978, sensitivity of 87.22%, and specificity of 96.02%. For the macular atrophy algorithm, the accuracy was 97.40% with an AUC of 0.982, sensitivity of 85.10%, and specificity of 98.34%. The CNV model displayed 91.14% accuracy with AUC of 0.881, sensitivity of 37.07, and specificity of 97.32%.

The META-PM CS’s total accuracy was 87.53%, with an accuracy of 92.08% for non-pathologic myopia (C0 or C1), 90.18% for diffuse atrophy (C2), 95.28% for patchy atrophy (C3), and 97.50% for macular atrophy (C4). The system’s accuracy decreased to 79.77% when researchers added the CNV model.

The system’s accuracy in distinguishing non-pathologic myopia from pathologic myopia was 92.08%. It had higher accuracy (87.53%) in categorizing myopic maculopathy than the trained grader (80.50%) and was close to the myopia specialist’s (89.00%). Its accuracy in identifying pathologic myopia was 78.06% in the open-source PALM dataset and 88.20% in the SEED dataset.

Limitations included the fact that all patients were Japanese and transfer learning may be required to apply findings to individuals of other ethnicities. Also, the diagnosis of the other “plus lesions” by fundus photographs alone was unreliable.

Reference

DU R, Xie S, Fang Y. Deep learning approach for automated detection of myopic maculopathy and pathologic myopia in fundus images. Ophthalmol Retina. Published online February 18, 2021. doi: 10.1016/j.oret.2021.02.006