AI-Model for Identifying Pathologic Myopia Based on Deep Learning Algorithms of Myopic Maculopathy Classification and “Plus” Lesion Detection in Fundus Images

Abstract

Background: Pathologic myopia (PM) associated with myopic maculopathy (MM) and “Plus” lesions is a major cause of irreversible visual impairment worldwide. Therefore, we aimed to develop a series of deep learning algorithms and artificial intelligence (AI)–models for automatic PM identification, MM classification, and “Plus” lesion detection based on retinal fundus images.Materials and Methods: Consecutive 37,659 retinal fundus images from 32,419 patients were collected. After excluding 5,649 ungradable images, a total dataset of 32,010 color retinal fundus images was manually graded for training and cross-validation according to the META-PM classification. We also retrospectively recruited 1,000 images from 732 patients from the three other hospitals in Zhejiang Province, serving as the external validation dataset. The area under the receiver operating characteristic curve (AUC), sensitivity, specificity, accuracy, and quadratic-weighted kappa score were calculated to evaluate the classification algorithms. The precision, recall, and F1-score were calculated to evaluate the object detection algorithms. The performance of all the algorithms was compared with the experts’ performance. To better understand the algorithms and clarify the direction of optimization, misclassification and visualization heatmap analyses were performed.Results: In five-fold cross-validation, algorithm I achieved robust performance, with accuracy = 97.36% (95% CI: 0.9697, 0.9775), AUC = 0.995 (95% CI: 0.9933, 0.9967), sensitivity = 93.92% (95% CI: 0.9333, 0.9451), and specificity = 98.19% (95% CI: 0.9787, 0.9852). The macro-AUC, accuracy, and quadratic-weighted kappa were 0.979, 96.74% (95% CI: 0.963, 0.9718), and 0.988 (95% CI: 0.986, 0.990) for algorithm II. Algorithm III achieved an accuracy of 0.9703 to 0.9941 for classifying the “Plus” lesions and an F1-score of 0.6855 to 0.8890 for detecting and localizing lesions. The performance metrics in external validation dataset were comparable to those of the experts and were slightly inferior to those of cross-validation.Conclusion: Our algorithms and AI-models were confirmed to achieve robust performance in real-world conditions. The application of our algorithms and AI-models has promise for facilitating clinical diagnosis and healthcare screening for PM on a large scale.

Development of a deep learning algorithm for myopic maculopathy classification based on OCT images using transfer learning

“…The model of OCT was the same as the one used in the training dataset. The protocol for the human-machine comparison was consistent with our previous study (15).…”

“…The present study demonstrates the potential application of the intelligent algorithms in the ophthalmic clinical tasks, especially for the highly myopic fundus lesions identification. Combined with our previous work (15,16), this work is also a part of our ongoing effort to develop the multiple-modal algorithms for automatic diagnosis of MM based on the multiple fundus examination methods including color fundus picture, OCT, OCTA, and FFA/ICGA, etc.…”

“…In this study, a total of two DL classification algorithms were trained to identify the targeted lesion categories, and the workflow was illustrated in Figure 2 . The OCT classification algorithm A, using the same deep residual network model architecture reported in our previous work ( 15 ), was pre-trained by ImageNet with no initialization parameters, and then the labeled OCT images were added to finalize the training. The OCT classification algorithm B was based on the classification algorithm developed in our previous study ( 15 ), using a multi-source cross-domain TL approach, sharing the prior distribution of model parameters, learning the feature representation of the OCT images, and further training to optimize the initialization parameters of the model.…”

“…In particular, all grades of macular retinoschisis, macular hole, macular hole with retinal detachment, and active/inactive myopic choroid neovascularization (mCNV) are easy to be distinguished with OCT images. In our previous work, DL showed comparable ability of ophthalmic fundus image classification and recognition when compared with the ordinary ophthalmologists, and even approached the level of retinal specialists (15). Meanwhile, transfer learning (TL) has been widely used in the development of medical DL algorithms, which means the infrastructure used is pre-trained in the huge ImageNet database, and then optimized with medical image data.…”

“…Combined with the automatic classification of MAM in our previous research, the full coverage of the intelligent automatic classification and diagnosis of MM using the ATN system was preliminary materialized. In addition, we also developed a new OCT classification algorithm using the TL approach based on the excellent classification algorithms (Algorithm I and Algorithm II) which was developed in our previous study ( 15 ).…”