Comparison of various neural network-based models for retinal lesion analysis

摘要

Identification and analysis of laser-induced lesions on the retina can be challenging in both the research and clinicalsettings depending on the age of a lesion and the imaging modality used for detection. Previous research exploringretinal damage thresholds utilized the consensus of an expert panel to confirm energies required for minimal visiblelesions, a method that includes some subjectivity. Because of this, there is a desire to develop an image processingarchitecture to accurately locate retinal laser lesions in images generated from clinically relevant modalities. Issues suchas imaging aberrations inducing circular artifacts, perceived stretch in lesions, and differences in the appearance oflesions across the dataset preclude use of traditional image processing tools. A database containing images of laserlesions has been developed in order to provide a reference for researchers and clinicians. In this work, we explored usingvarious Convolutional Neural Network (CNN) architectures and preprocessing techniques to more objectively identifyand analyze retinal laser lesions. Specifically, we developed frequency domain filtering techniques in order to emphasizelesion qualities. We consider this task to be one of image segmentation to make the networks somewhat size invariant.Since the lesions account for a small amount of the image pixels, we implemented an intersection-based loss function.We evaluated the performance of our trained networks against more complicated architecture variants. Additionally, wetrained a network to segment and classify lesions as the result of photochemical, photomechanical or photothermaldamage.