CXR ‐ ODDet : An Omni‐Decoupled Multi‐Class Lesion Localization Framework for Automatic Chest X‐Ray Analysis

ABSTRACT Chest X‐ray (CXR) is a typical radiological examination for screening thoracic diseases. Existing deep detection networks have achieved great success in computer‐aid diagnosis for single diseases but are challenged by the limited universality of multiple diseases. To address this issue, we propose an omni‐decoupled Framework for multi‐class abnormalities localization in CXR, called CXR‐ODDet. Specifically, we first devise a Granularity‐Decoupled Encoder (GD‐Encoder) that leverages the integration of lesion‐common and lesion‐specific knowledge to generate discriminative feature representations across different disease types. Moreover, we introduce a Task‐Decoupled Encoder (TD‐Decoder) to prevent the different lesion knowledge from interfering with each other by providing a private decoding workflow for each lesion. We take the GD‐Encoder and TD‐Decoder to build our CXR‐ODDet framework. Extensive experiments prove that CXR‐ODDet outperforms state‐of‐the‐arts in multi‐class abnormality localization of chest X‐rays. In particular, the proposed CXR‐ODDet reveals a significant Average Precision (AP 0.4 }_{0.4 improvement of 3.4% and 4.2% in two CXR datasets.