Symmetry-Aware Siamese Network: Exploiting Pathological Asymmetry for Chest X-Ray Analysis

The human body shows elements of bilateral symmetry for various body parts, including the lung. This symmetry can be disturbed by a variety of diseases or abnormalities, e.g. by lung diseases such as pneumonia. While radiologists use lung field symmetry information in their radiological examinations to analyze chest X-rays, it is still underutilized in the field of computer vision. To investigate the potential of pathologically induced asymmetry of the lung field for the automatic detection of healthy and diseased patients, we implement a symmetry-aware architecture. The model is based on a Siamese network with a DenseNet backbone and a symmetry-aware contrastive loss function. Two different processing pipelines are investigated: first, the scan is processed as a whole image, and second, the left and right lung fields are separated. This enables an independent determination of the most important features of each lung field. Compared to state-of-the-art baseline models (DenseNet, Mask R-CNN), symmetry-aware training can improve the AUROC score by up to 10%. Furthermore, the findings indicate that, by integrating the bilateral symmetry of the lung field, the interpretability of the models increases. The generated probability maps show a stronger focus on lung field and disease features compared to state-of-the-art algorithms like Grad-Cam++ for heat map generation or Mask R-CNN for object detection.