CDC-Net: Cascaded decoupled convolutional network for lesion-assisted detection and grading of retinopathy using optical coherence tomography (OCT) scans. (September 2021)

• Record Type: Journal Article
• Authors: Hassan, Bilal; Qin, Shiyin; Hassan, Taimur; Akram, Muhammad Usman; Ahmed, Ramsha; Werghi, Naoufel
• Is Part Of: Biomedical signal processing and control. Volume 70(2021)
• Publication Date: 2021-09
• Subjects: Optical coherence tomography (OCT) -- Deep learning -- Lesion segmentation -- Retinopathy grading -- Retinal analysis; Signal processing -- Periodicals; Biomedical engineering -- Periodicals; Signal Processing, Computer-Assisted -- Periodicals; Image Processing, Computer-Assisted -- Periodicals; Biomedical Engineering -- Periodicals; 610.28
• DOI: 10.1016/j.bspc.2021.103030
• Languages: English
• ISSNs: 1746-8094

Abstract

Abstract: Retinopathy refers to any injury in the retinal region of the eye that can lead to distorted vision or even blindness. The segmentation of retinal lesions or biomarkers is crucial for the precise classification and grading of retinopathy. Optical coherence tomography imaging is the widely used eye examination tool by ophthalmologists due to its comprehensive visualization of the retinal lesions, which can assist in the prompt treatment of retinal conditions. However, due to vast clinical optical coherence tomography applications and the incidence of ocular syndromes, the number of scans collected daily outweighs ophthalmologists' capacity to interpret these in a meaningful way. Many studies have been proposed previously to address this issue using optical coherence tomography scans. However, to the best of our knowledge, no framework can perform joint segmentation of the retinal lesions and retinopathy grading. In this paper, we propose a novel framework to address this shortcoming. We propose a new cascaded decoupled convolutional network comprising two separate modules that work together to perform lesion-assisted grading of retinopathy according to the clinical standards. We thoroughly evaluated the proposed framework using 26841 multi-vendor scans spanned over four publicly accessible datasets. The obtained results validate the efficacy of the proposed scheme over other state-of-the-art frameworks, where it achieved the mean Dice score of 0.820 (3.66% improvement) in segmentation of retinal lesions and 98.89% accuracy in the grading of retinopathy with 98.34% true positive rate and 99.17% true negative rate. Highlights: CDC-Net gives intuitive analysis of OCT scans w.r.t. lesions and underlying pathology. It integrates a comprehensive preprocessing mechanism to boost the performance. CDC-Net is a novel model to simultaneously segment four lesions (IRF, SRF, PED, RPD). Using cascaded decoupled CNN modules, it learns inherent multiscale OCT features. CDC-Net outperforms other SOTA schemes in lesion segmentation & retinopathy grading.