Differential diagnosis between polypoidal choroidal vasculopathy (PCV) and Age-related macular degeneration (AMD) using Deep Neural Network

Abstract

Purpose : Polypoidal choroidal vasculopathy (PCV) is increasingly recognized as an important cause of exudative maculopathy especially in Asian polulation, which often get misdiagnosed as wet age-related macular degeneration (AMD) due to similar pathology therefore affect the effective treatment to be applied. Indocyanine green angiography (ICGA) is the investigation used for definitive diagnosis of PCV despite its limited availability and invasive nature. On the other hand, PCV exhibit characteristic features in optical coherent tomography (OCT). In this study, we explored the feasibility of using the deep-learning technique to detect characteristic imaging-based features using only non-invasive structural volumetric OCT imaging, and achieve automatic differential diagnosis between PCV and AMD.

Methods : A hospital-based cross-sectional study was done in a tertiary eye care center in South India. Electronic medical records (EMR) were reviewed of patients who reported to the Vitreoretinal services. Patients with a clinical suspicion of PCV underwent fundus flurescein angiography (FA), ICG angiography structural OCT. The diagnosis of PCV was done based on the EVEREST criteria by 3 senior retina specialists independently. PCV patients and 34 AMD patients were included in this study.

All OCT were firstly extracted into individual and B-scans from PCV and AMD patients respectively. Deep residual neural network along with transfer learning were used to train automatic disease classification. 10 subjects in each group were hold-out for testing, with remaining data splited into for training (PCV:AMD=2224:2165) and validation (PCV:AMD608/604). The diagnosis for each patient in the testing set is determined by aggregating the classification results from each individual B-scan. Patients with more than 50% of B-scans classified as a specific disease (PCV or AMD) were determined as the predicted diagnosis.

Results : 9 out of 10 patients in both PCV and AMD group in the test patients were correctly diagnosed. The Class Activation Map were also generated to provide visual explainable model behavior to facilitate clinical translation.

Conclusions : In this study, we demonstrated the feasibility of using non-invasive structural OCT imaging technique to differentiate patient with PCV from AMD with great diagnosis accuracy.

This is a 2020 ARVO Annual Meeting abstract.

 

Figure 1

 

Figure 2