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

Abstract

PurposeTo apply deep learning (DL) techniques to develop an automatic intelligent classification system identifying the specific types of myopic maculopathy (MM) based on macular optical coherence tomography (OCT) images using transfer learning (TL).MethodIn this retrospective study, a total of 3,945 macular OCT images from 2,866 myopic patients were recruited from the ophthalmic outpatients of three hospitals. After culling out 545 images with poor quality, a dataset containing 3,400 macular OCT images was manually classified according to the ATN system, containing four types of MM with high OCT diagnostic values. Two DL classification algorithms were trained to identify the targeted lesion categories: Algorithm A was trained from scratch, and algorithm B using the TL approach initiated from the classification algorithm developed in our previous study. After comparing the training process, the algorithm with better performance was tested and validated. The performance of the classification algorithm in the test and validation sets was evaluated using metrics including sensitivity, specificity, accuracy, quadratic-weighted kappa score, and the area under the receiver operating characteristic curve (AUC). Moreover, the human-machine comparison was conducted. To better evaluate the algorithm and clarify the optimization direction, the dimensionality reduction analysis and heat map analysis were also used to visually analyze the algorithm.ResultsAlgorithm B showed better performance in the training process. In the test set, the algorithm B achieved relatively robust performance with macro AUC, accuracy, and quadratic-weighted kappa of 0.986, 96.04% (95% CI: 0.951, 0.969), and 0.940 (95% CI: 0.909–0.971), respectively. In the external validation set, the performance of algorithm B was slightly inferior to that in the test set. In human-machine comparison test, the algorithm indicators were inferior to the retinal specialists but were the same as the ordinary ophthalmologists. In addition, dimensionality reduction visualization and heatmap visualization analysis showed excellent performance of the algorithm.ConclusionOur macular OCT image classification algorithm developed using the TL approach exhibited excellent performance. The automatic diagnosis system for macular OCT images of MM based on DL showed potential application prospects.

Performances of artificial intelligence in detecting pathologic myopia: a systematic review and meta-analysis

“…Figure 1 shows the flowchart of the literature eligibility process. Finally, 22 studies were included for systematic review [ 12 – 33 ], and 14 of them were included for quantitative meta-analysis [ 12 , 13 , 16 , 19 – 23 , 26 , 27 , 29 – 31 , 33 ].

Fig.

…”

“…For the detection of PM-CNV, the forest plots for the included studies and the SROC curve plot are shown in Fig. 3 [ 12 , 13 , 21 , 26 , 33 ]. The summary AUC was 0.99 (95% CI: 0.97 to 0.99), and the pooled SEN, SPE, PLR, NLR, and DOR were 0.94 (95% CI: 0.90 to 0.97), 0.96 (95% CI: 0.94 to 0.98), 25.9 (95% CI: 16.1 to 41.7), 0.06 (95% CI: 0.03 to 0.10), and 435 (95% CI: 220 to 860), respectively.…”

“…We assessed the quality of all included studies using the QUADAS-2 tool, and the results are presented in Supplementary Appendix 3. 7 studies (31.8%) were graded as having a low risk of bias in all 4 domains [16,18,20,[26][27][28]33]. 12 studies (54.5%) had at least one domain with applicability concerns [14,15,17,19,21,22,24,25,[29][30][31][32].…”

Performances of artificial intelligence in detecting pathologic myopia: a systematic review and meta-analysis

“…Finally, 22 studies were included for systematic review, and 14 of them were included for quantitative meta-analysis. [12,13,16,[19][20][21][22][23]26,27,[29][30][31]33] The characteristics of all eligible studies are summarized in table 1. In total, 348,861 fundus images and 22,560 OCT images were used for training, testing and validation.…”

“…For the detection of PM-CNV, the forest plots for the included studies and the SROC curve plot are shown in gure 3. [12,13,21,26,33] The summary AUC was 0.99 (95% CI 0.97 to 0.99), and the pooled SEN, SPE, PLR, NLR, and DOR were 0.94 (95% CI 0.90 to 0.97), 0.96 (95% CI 0.94 to 0.98), 25.9 (95% CI 16.1 to 41.7), 0.06 (95% CI 0.03 to 0.10), and 435 (95% CI 220 to 860), respectively.…”