Tuberculosis lesions in CT images inferred using 3D-CNN and multi-task learning

Tuberculosis (TB) is an infectious disease that very frequently damage the lungs and that has a high incidence and mortality rate. The longitudinal assessment of pulmonary affectation is a clear need to boost the development of novel drugs and to control the spread of the disease. In this manuscript, we train a computational model able to infer TB manifestations present in each lung lobe of x-ray Computer Tomography (CT) scans by employing the associated radiologist reports as ground truth. We do so instead of using the classical manually delimited segmentation masks. The proposed learning strategy adapts the V-Net model, which allow us to employ full 3D volumes in order to obtain fine grain features. The model is successfully optimized by a novel loss function for multi-task learning. The loss function employs the model uncertainty to weight the regression and binary tasks. Our results are promising with a Root Mean Square Error of 1.14 in the number of nodules and F¹-scores above 0.85 for the most prevalent TB lesions (i.e., Conglomerations, cavitations, consolidations, trees in bud) when considering the whole lung.