3D-CNN in Drug Resistance Detection and Tuberculosis Classification

Object classification is a very demanding field in computer vision, especially when dealing with medical imaging datasets, which are often small and have unbalanced distributions. Deep learning (DL) methods have proven to be effective in dealing with such problems and have established themselves as the state-of-the-art. ImageCLEFtuberculosis is a challenge that encompasses the classification problem on medical images, and is divided into two subtasks: Drug Resistance Detection and Tuberculosis classification. For both subtasks, provided images were pre-processed to segment the lungs from the CT volumes. Afterwards, pre-processed CT volumes were fed in batches to a 3D convolutional neural network. Test results for the Drug Resistance detection task scored an accuracy of 46.5% and AUC of 0.46, while in the Tuberculosis classification task an accuracy of 24% and Cohen’s Kappa value of 0.022 were obtained. Using data augmentation and weight normalization, the overfitting problem could be reduced, and submitted models’ performance improved.