Nearest Neighbor Radiomics for Self-Supervised Chest X-ray Pneumonia Identification

Cailin Winston, Caleb Winston, Chloe Winston

Keywords: Constrastive Learning, Self-Supervised Learning, Radiomics, Chest X-Ray, Pneumonia Identification

TL;DR: Self-supervised training for chest X-ray models using positive pairs that have nearest neighboring radiomics features can boost accuracy.

Abstract: Self-supervised training minimizes a contrastive loss objective for unlabeled data. Contrastive loss estimates the distance in the latent space between positive pairs, which are pairs of images that are expected to have the same label. For medical images, choosing positive pairs is challenging because simple transformations like rotations or blurs are not class-invariant. In this paper, we show that choosing positive pairs with nearest-neighbor radiomics features for self-supervised training improves chest X-ray pneumonia identification accuracy by 8.4% without labeled data.