EGFDA: Experience-guided Fine-grained Domain Adaptation for cross-domain pneumonia diagnosis

Abstract

Although recent advances in deep learning have led to accurate pneumonia diagnoses, their heavy reliance on data annotation hinders their expected performance in clinical practice. Unsupervised domain adaptation (UDA) methods have been developed to address the scarcity of annotations. Nevertheless, the diverse manifestations of pneumonia pose challenges for current UDA methods, including spatial lesion-preference bias and discriminative class-preference bias. To overcome these problems, we propose an Experience-Guided Fine-grained Domain Adaptation (EGFDA) framework for automatic cross-domain pneumonia diagnosis. Our framework consists of two main modules: (1) Gradient-aware Lesion Area Matching (GaLAM), which aims to reduce the global domain gap while avoiding misleading from lesion-unrelated targets, and (2) Reweighing Smooth Certainty-aware Matching (RSCaM), which aims to match class space with a smooth certainty-aware feature mapping to guide the model to learn more precise class-discriminative features. Benefiting from the collaboration between GaLAM and RSCaM, the proposed EGFDA is able to process unlabeled samples following a pattern similar to the diagnostic experience of physicians, that is, first locating the disease-related lesion area and then performing fine-grained discrimination. Comprehensive experiments on three different tasks using six datasets demonstrate the superior performance of our EGFDA. Furthermore, extensive ablation studies and visual analyses highlight the remarkable interpretability and generalization of the proposed method.