Hayagreev Jeyandran
Research Mentor: Matthias Wilms
Mentor Department: Department of Radiology, Medicine
Author(s): Hayagreev Jeyandran, Matthias Wilms
Session: Session 3 (11:00 AM – 11:50 AM)
Presentation Type: Poster 51
Abstract
Color fundus photography (CFP) is a low-cost, highly scalable 2D retinal imaging modality widely used in ophthalmology to screen for different eye diseases, such as diabetic retinopathy, glaucoma, and ocular tumors. Recent years have seen a proliferation of deep learning techniques for the high-throughput detection of early-stage eye diseases from CFPs to support clinicians. These models, however, learn from training distributions with fixed resolutions (e.g., 224 x 224). Therefore, they are constrained to utilizing fixed-resolution input images during inference. This is achieved by downsampling the original data acquired at different camera-specific resolutions, often containing several thousand pixels in each dimension, which removes and masks subtle features such as small-vessel anomalies and early lesions. In response, we propose a fully resolution-agnostic pipeline which utilizes Implicit Neural Representations (INRs) to continuously parametrize discrete CFPs as the weights of a deep neural network. When the INR weights are encoded in a low-dimensional latent space, we show that it is possible to classify the original CFP into one of five ocular conditions and a normal class. We evaluate this architecture against state-of-the-art, fixed-resolution networks over a distribution of 2,300 testing images, with the aim of demonstrating improved classification robustness by preserving high-frequency spatial features. When integrated with techniques from interpretable machine learning, our model can enable advancements in clinical decision support for ophthalmologists and benefit patient care in the early detection of eye conditions before vision impairment starts.
