Learning to Generalize towards Unseen Domains via a Content-Aware Style Invariant Model for Disease Detection from Chest X-rays

Performance degradation due to distribution discrepancy is a longstandingchallenge in intelligent imaging, particularly for chest X-rays (CXRs). Recentstudies have demonstrated that CNNs are biased toward styles (e.g.,uninformative textures) rather than content (e.g., shape), in stark contrast tothe human vision system. Radiologists tend to learn visual cues from CXRs andthus perform well across multiple domains. Motivated by this, we employ thenovel on-the-fly style randomization modules at both image (SRM-IL) and feature(SRM-FL) levels to create rich style perturbed features while keeping thecontent intact for robust cross-domain performance. Previous methods simulateunseen domains by constructing new styles via interpolation or swapping stylesfrom existing data, limiting them to available source domains during training.However, SRM-IL samples the style statistics from the possible value range of aCXR image instead of the training data to achieve more diversifiedaugmentations. Moreover, we utilize pixel-wise learnable parameters in theSRM-FL compared to pre-defined channel-wise mean and standard deviations asstyle embeddings for capturing more representative style features.Additionally, we leverage consistency regularizations on global semanticfeatures and predictive distributions from with and without style-perturbedversions of the same CXR to tweak the model's sensitivity toward contentmarkers for accurate predictions. Our proposed method, trained on CheXpert andMIMIC-CXR datasets, achieves 77.32$\pm$0.35, 88.38$\pm$0.19, 82.63$\pm$0.13AUCs(%) on the unseen domain test datasets, i.e., BRAX, VinDr-CXR, and NIHchest X-ray14, respectively, compared to 75.56$\pm$0.80, 87.57$\pm$0.46,82.07$\pm$0.19 from state-of-the-art models on five-fold cross-validation withstatistically significant results in thoracic disease classification.