The authors study thorax disease classification on radiographic images. Effective extraction of features for the disease areas is crucial for disease classification, but existing methods struggle to handle the adverse effect of noise and background on radiographic images.
To address this challenge, the authors propose a novel Low-Rank Feature Learning (LRFL) method. LRFL is motivated by the observation that the low-rank projection of the ground truth class labels possesses the majority of the information. It adds a truncated nuclear norm regularization term to the training loss to promote low-rank features, which helps discard high-frequency features related to noise and background.
The authors provide a theoretical result on the sharp generalization bound for LRFL, justifying the benefits of low-rank learning. Extensive experiments on three thorax disease datasets (NIH-ChestX-ray, COVIDx, and CheXpert) demonstrate that LRFL outperforms state-of-the-art methods, achieving new record performance on all three datasets. The authors also show that LRFL is particularly effective in small data regimes, exhibiting strong generalization capabilities.
In un'altra lingua
dal contenuto originale
arxiv.org
Approfondimenti chiave tratti da
by Rajeev Goel,... alle arxiv.org 05-01-2024
https://arxiv.org/pdf/2404.18933.pdfDomande più approfondite