Pathological myopia classification with simultaneous lesion segmentation using deep learning

Abstract

Purpose: To assess the use of deep learning for detection of pathological myopia (PM) and semantic segmentation of myopia-related lesions from fundus images. Methods: This investigation reports on the results of deep learning models developed for the recently introduced Pathological Myopia (PALM) dataset, which consists of 1200 images. Evaluation metrics include area under the receiver operating characteristic curve (AUC) for PM detection, Euclidean distance for fovea localization, and Dice and F1 metrics for the semantic segmentation tasks (optic disc, retinal atrophy and retinal detachment). We also introduce a new Optic Nerve Head (ONH)-based prediction enhancement for the segmentation of atrophy and fovea. Results: Models trained with 400 available training images achieved an AUC of 0.9867 for PM detection, and a Euclidean distance of 58.27 pixels on the fovea localization task, evaluated on a test set of 400 images. Dice and F1 metrics for semantic segmentation of lesions scored 0.9303 and 0.9869 on optic disc, 0.8001 and 0.9135 on retinal atrophy, and 0.8073 and 0.7059 on retinal detachment, respectively. Conclusions: We report a successful approach for a simultaneous classification of pathological myopia and segmentation of associated lesions. Our work was acknowledged with an award in the context of the "Pathological Myopia detection from retinal images" challenge held during the IEEE International Symposium on Biomedical Imaging (April 2019). Considering that (pathological) myopia cases are often identified as false positives and negatives in glaucoma deep learning models, we envision that the current work could aid in future research to discriminate between glaucomatous and highly-myopic eyes, complemented by the localization and segmentation of landmarks such as fovea, optic disc and atrophy.

An Artificial Intelligent Risk Classification Method of High Myopia Based on Fundus Images

“…In recent years, with the rapid development of the artificial intelligence technology and deep learning methods, many researchers have applied them to various image processing problems. New techniques and methods that analyze fundus images for high myopia have been continuously emerging [ 5 , 6 , 7 , 8 , 9 , 10 ]. These methods use computer-aided technologies to automatically analyze and diagnose lesions associated with high myopia in the absence of experienced ophthalmologists and professional optometry instruments.…”

“…The first task of the 2019 PALM Color Fundus Photographic Pathological Myopia Challenge [ 8 ] was the qualitative classification of pathological myopia. Participating teams used different CNN models [ 9 , 10 , 11 ] to predict the risk of pathological myopia using fundus images. However, these classifications only distinguished pathological myopia from nonpathological myopia, and certain fundus images classified as nonpathological myopia still have a certain high myopia risk (such as “simple high myopia” defined in [ 3 ]).…”

Predicting Optical Coherence Tomography-Derived High Myopia Grades From Fundus Photographs Using Deep Learning

“…Automatic diagnosis of retinal detachment and/or PMCNV by fundus photographs has been reported in some studies ( 25 , 26 ). Hemelings et al developed an algorithm for PM detection using 400 fundus photographs, and the F1 score for retinal detachment was 0.71 ( 25 ).…”

“…Some researchers have performed automatic identification and segmentation of myopic lesions based on fundus photographs, and achieved promising performance ( 25 – 28 ). Tan et al developed an algorithm which achieved high diagnostic performance, with an AUC of 0.913 for high myopia and 0.969 for myopic maculopathy ( 26 ).…”

Applications of Artificial Intelligence in Myopia: Current and Future Directions

“…By analyzing demographic and clinical information, retinal fundus photograph data and genotyping data from 2,258 subjects, this method achieved an AUC of 0.888 and outperformed the detection results obtained from the use of demographic and clinical information (an AUC of 0.607), imaging data (an AUC of 0.852) or genotyping data (an AUC of 0.774) alone, with increases of 46.3%, p < 0.005; 4.2%, p = 0.19; and 14.7%, P < 0.005, respectively. Recently, Hemelings et al ( 39 ) developed a successful approach based on a DL algorithm for the simultaneous detection of PM, with an AUC of 0.9867, and the segmentation of myopia-induced lesions. Other similar studies have also been reported, such as those identifying the different types of lesions of myopic maculopathy automatically from fundus photographs with DL models ( 40 , 41 ).…”