Detection of Respiratory Disease Patterns Using Mask R-CNN

The analysis and identification of pathological signs associated with different respiratory diseases are is not an easy task. One of the imaging modalities for these signs identification is examining chest CT scans. However, it requires expert knowledge to avoid human error. The purpose of this work is to implement, test, and analyze the performance of a neural network based on a mask R-CNN model able to identify some pathological signs of respiratory disease. The CT images used were manually labeled and pre-classified as positive and negative cases by specialists to prepare them for the training process. Preliminary results reached detection of ground-glass opacity with a sensitivity of 81.89% using the validation set and 92.66% using the test set. Nevertheless, low percentages were obtained for pulmonary nodules detection with a sensitivity of 51.08 and 40.34% using validation and test sets, respectively.