A0547
Title: ImgKnock: Knockoff inference with fundus images for glaucoma diagnosis
Authors: Zhe Fei - UC Riverside (United States) [presenting]
Abstract: ImgKnock is an innovative pipeline that leverages knockoff inference and deep learning for the analysis of optic disc images in glaucoma. Knockoff inference, known for its ability to control false discovery rates in high-dimensional data, is adapted here to handle high-resolution medical images. ImgKnock uniquely generates latent knockoff features of fundus photographs, enabling the prediction of glaucoma-related outcomes and the identification and testing of important image features for accurate diagnosis. The method extends knockoff generation to non-tabular data, maintaining key features like the swapping property and ensuring robust feature selection. The pipeline consists of the following: i) Latent feature learning from the image data; ii) Knockoff generation on the latent features; iii) Knockoff feature selection with FDR control; i) Interpreting important features for glaucoma detection. This approach facilitates precise inference of feature importance, which is crucial for understanding the disease. ImgKnock has been successfully applied to various image datasets, including MNIST and CIFAR-10, demonstrating its versatility. ImgKnock's potential to contribute to improved diagnostic methods and a deeper understanding of the disease is highlighted.
