This research presents an ordinal regression framework for assessing disease severity in chest radiographs using deep learning. The proposed framework divides the ordinal regression problem into three parts: a model, a target function, and a classification function. Various encoding methods, including one-hot, Gaussian, progress-bar, and soft-progress-bar are used with ResNet50 and ViT-B-16 deep learning models. The study finds that the choice of encoding has a significant impact on performance, and the best encoding depends on the chosen weighting of Cohen’s kappa and the model architecture used. The researchers used their own dataset of 193k images, labeled by experienced radiologists on a graded scale, to validate their work.
Publication date: 8 Feb 2024
Project Page: https://arxiv.org/abs/2402.05685v1
Paper: https://arxiv.org/pdf/2402.05685
