Deep Learning Models for Tuberculosis Detection from Chest X-ray Images

This paper explores the usefulness of transfer learning on medical imaging for tuberculosis detection. We show an improved method for transfer learning over the regular method of using ImageNet weights. We also discover that the low-level features from ImageNet weights are not useful for imaging tasks for modalities like X-rays and also propose a new method for obtaining low level features by training the models in a multiclass multilabel scenario. This results in an improved performance in the classification of tuberculosis as opposed to training from a randomly initialized settings. In other words, we have proposed a better way for training in a data constrained setting such as the healthcare sector.