From 3D to 2D: Transferring knowledge for rib segmentation in chest X-rays

Abstract

Chest X-rays are the most common type of biomedical radiologic exam, being widely adopted for the diagnosis of a myriad of illnesses in the thoracic region. Computed Tomography – even though being more expensive and rare – is also a useful tool for the detection of several illnesses and surgery planning, providing volumetric information. This paper proposes a methodology aiming to leverage the larger amounts of spatial information and lack of occlusion in tomographic images to aid in the rib segmentation of 2D X-ray images by means of Domain Adaptation. We perform extensive quantitative and qualitative experiments to test the capabilities of this methodology in segmenting ribs in 7 X-ray datasets with distinct visual features, using 6 different metrics and without any use of rib segmentation labels from the target image sets. In order to encourage reproducibility, all data and code used in this research is publicly available online, including a new 2D Digitally Reconstructed Radiograph generated from tomographic data and a new pixel-level label map for the JSRT Chest X-ray dataset. We also publicize our generalizable pretrained models for both rib segmentation in Chest X-rays and lung field segmentation in Digitally Reconstructed Radiographs. Results show that the proposed pipeline outperforms shallow rib segmentation baselines in almost all quantitative metrics and produce higher fidelity pixel-map predictions than simply using the pretrained Neural Networks on the flattened 3D data, mainly in datasets where domain shift is more pronounced. The use of Conditional Domain Adaptation also allows the method to perform inference on all 7 X-ray datasets using one single model, achieving over 0.856 of AUC on OpenIST and 0.934 of AUC on JSRT, with Dice scores of 0.68 and 0.69 in these two datasets, respectively.