Thorax Disease Classification Using Low-Rank Feature Learning
Abstract: In the field of medical imaging, deep neural networks such as Convolutional Neural Networks (CNNs) and Vision Transformers (ViT) have demonstrated remarkable achievements. In this paper, we focus on classifying thorax diseases based on radiographic images. The key to the success of classification involves effectively extracting features from disease-impacted areas in radiographic images. Although various neural network architectures and training methods, including self-supervised learning through contrastive/restorative techniques, have been utilized for such classification tasks, there remains a lack of systematic approaches to mitigate the negative impacts of noise and non-disease elements in the images. To tackle this issue, we introduce a new Low-Rank Feature Learning (LRFL) technique in this study, which can be implemented in the training processes of different neural networks. The LRFL approach is both empirically inspired by a Low Frequency Property (LFP) and theoretically supported by our precise generalization bounds for neural networks using low-rank features. Notably, LFP is prevalent not only in deep neural networks across general machine learning applications but also across all thorax medical datasets examined in this study. In our empirical evaluation, the LRFL method, when applied to a ViT or CNN that has been pre-trained on unlabeled chest X-rays using Masked Autoencoders (MAE), outperforms existing methods in terms of multi-class area under the receiver operating curve (mAUC) and classification accuracy. The code is available at \url{https://anonymous.4open.science/r/medical_projects-BBFE/}.
Submission Length: Regular submission (no more than 12 pages of main content)
Assigned Action Editor: ~Efstratios_Gavves1
Submission Number: 3170
Loading