ImgKnock: Novel Knockoff Generation and Feature Selection for Image Data via Latent Representations

Music City Center 

Over 3 million Americans currently have glaucoma, a series of eye conditions that damage the optic nerve, leading to more severe eye issues. Diagnosis and monitoring of glaucoma can be accomplished through examination of fundus images, such as thinning of the neuroretinal rim. While traditional feature selection techniques can be applied to pixelated fundus image data, they often struggle with high dimensionality, computational inefficiency, and procedural rigidity. To resolve these issues and control FDR, we present a novel approach that leverages latent representation learning to construct higher-level features from image data and generate knockoffs of the latent features, followed by knockoff feature selection with FDR control. Called ImgKnock, our four-step procedure uses a deep latent representation learning-based approach integrated with a model-X knockoffs framework. Simulations are conducted using the common MNIST and CIFAR-10 datasets to demonstrate the efficacy of ImgKnock. Results indicate proper FDR control, particularly with MNIST data, showing an AUC metric of up to 0.889. The proposed ImgKnock is also applied to fundus images from the UCLA Stein Eye Institute.

knockoff selection

latent representation learning

FDR control

fundus images

self-supervised learning 

Section on Statistics in Imaging