Abstract
Purpose :
Retinal vascular morphology provides valuable information for both ophthalmic disease and systemic disease (termed 'oculomics'), e.g., atherosclerosis and diabetes mellitus, in a rapid and non-invasive way. To help recognise high risk cases of ophthalmic and systemic disease through observing the changes of retinal vascular morphology, we propose a deep learning pipeline to automatically analyse the vascular morphology (AutoMorph) which measures 12 kinds of metrics, such as vessel calibre and tortuosity.
Methods :
AutoMorph consists of four functional modules: image pre-processing, image quality grading, anatomical segmentation, including binary vessel, artery/vein, and optic disc/cup segmentation, and vascular morphology feature measurement. Image quality grading and anatomical segmentation use the most recent deep learning techniques and are trained on 12 public datasets, such as DRIVE, STARE, and CHASEDB1. We employ model ensemble strategy to achieve robust results and analyse the prediction confidence to rectify false gradable cases in image quality grading. We quantitatively validate module performance on 6 external publicly available datasets including EyePACS image quality dataset (EyePACS-Q), DDR, AV-WIDE, DR-HAGIS, IOSTAR-AV, and IDRID datasets.
Results :
The EfficientNet-b4 architecture used in the image grading module achieves comparable performance to the state-of-the-art method in EyePACS-Q, with an F1-score of 0.86. The confidence analysis reduces 76% of false gradable images. The binary vessel segmentation module achieves an F1-score of 0.73 on AV-WIDE and 0.78 on DR-HAGIS. The artery/vein module scores 0.66 on IOSTAR-AV, and optic disc/cup module achieves 0.94 in disc segmentation in IDRID. These results verify that AutoMorph performs well in external validation, being quantitatively on par or even better than some recent work in internal validation.
Conclusions :
AutoMorph performs well even when the external validation data shows significant difference to the training data, e.g., validation on ultra-wide field retinal photography. The fully automatic pipeline integrates recent technical work to facilitate 'oculomics' research.
This abstract was presented at the 2022 ARVO Annual Meeting, held in Denver, CO, May 1-4, 2022, and virtually.