Improving Tuberculosis Recognition on Bone-Suppressed Chest X-Rays Guided by Task-Specific Features

Abstract

Early tuberculosis (TB) screening through chest X-rays (CXRs) is essential for timely detection and treatment of the disease. Previous studies have shown that texture abnormalities in CXRs can be enhanced by bone suppression techniques, which may potentially improve TB diagnosis performance. However, existing TB datasets with CXRs usually lack bone-suppressed images, making it difficult to take advantage of bone suppression for TB recognition. Also, existing bone suppression models are usually trained on a relatively small dual-energy subtraction (DES) dataset with CXRs, without considering the image specificity of TB patients. To this end, we propose a bone-suppressed CXR-based tuberculosis recognition (BCTR) framework, where diagnosis-specific deep features are extracted and used to guide a bone suppression (BS) model to generate bone-suppressed CXRs for TB diagnosis. Specifically, the BCTR consists of a classification model for TB diagnosis and an image synthesis model for bone suppression of CXRs. The classification model is first trained on original CXRs from multiple TB datasets such as the large-scale TBX11K. The image synthesis model is trained on a DES dataset to produce bone-suppressed CXRs. Considering the heterogeneity of CXR images from the TB datasets and the DES dataset, we proposed to extract multi-scale task-specific features from the trained classification model and transfer them (via channel-wise addition) to the corresponding layers in the image synthesis model to explicitly guide the bone suppression process. With the bone-suppressed CXRs as input, the classification model is further trained for multi-class TB diagnosis. Experimental results on five TB databases and a DES dataset suggest that our BCTR outperforms previous state-of-the-arts in automated tuberculosis diagnosis.