Prediction of Multidrug-Resistant TB from CT Pulmonary Images Based on Deep Learning Techniques

While tuberculosis (TB) disease was discovered more than a century ago, it has not been eradicated yet. Quite contrary, at present, TB constitutes one of the top 10 causes of death and has shown signs of increasing. To complement the conventional diagnostic procedure of applying microbiological culture that takes several weeks and remains expensive, high resolution computer tomography (CT) of pulmonary images has been resorted to not only for aiding clinicians to expedite the process of diagnosis but also for monitoring prognosis when administering antibiotic drugs. This research undertakes the investigation of predicting multidrug-resistant (MDR) patients from drug-sensitive (DS) ones based on CT lung images to monitor the effectiveness of treatment. To contend with smaller data sets (i.e., hundreds) and the characteristics of CT TB images with limited regions capturing abnormities, patch-based deep convolutional neural network (CNN) allied to support vector machine (SVM) classifier is implemented on a collection of data sets from 230 patients obtained from the ImageCLEF 2017 competition. As a result, the proposed architecture of CNN + SVM + patch performs the best with classification accuracy rate at 91.11% (79.80% in terms of patches). In addition, a hand-crafted SIFT based approach accomplishes 88.88% in terms of subject and 83.56% with reference to patches, the highest in this study, which can be explained away by the fact that the data sets are in small numbers. Significantly, during the Tuberculosis Competition at ImageCLEF 2017, the authors took part in the task of classification of 5 types of TB disease and achieved the top one with regard to averaged classification accuracy (i.e., ACC = 0.4067), which is also premised on the approach of CNN + SVM + patch. On the other hand, when the whole slices of 3D TB data sets are applied to train a CNN network, the best result is achieved through the application of CNN coupled with orderless pooling and SVM at 64.71% accuracy rate.