Instance Segmentation of Chest X-Ray Images Using Mask RCNN for Automated Pulmonary Tuberculosis Detection

One of the most dangerous infectious respiratory diseases that affect people is tuberculosis (TB). It is essential to detect TB disease in the early stage; otherwise, it can lead to a life-threatening problem. This study proposed an automatic TB detection method based on deep learning, in which automatic analysis and classification of Chest X-Ray images (CXR) for Tb or not Tb was performed. With the Transfer Learning method for automated TB detection that reduces the cost and time, a deep learning-based approach called Mask RCNN is used for semantic interpretation of images. It is a state-of-the-art model trained to determine pixel level/instance segmentation for each object in the image used for TB detection with generalized feature extraction. ResNet 50 and ResNet 101 models were used for the experimental results and to compare the classification performance with DenseNet and Inception models. The experimental results indicate that Resnet 101 produces the best result with a Mean Average Precision of 94.33, and ResNet-50 produces 89 with a learning rate of 0.001.