ToraxIA: Virtual Assistant for Radiologists Based on Deep Learning from Chest X-Ray

Abstract

Misdiagnosis of pulmonary pathology may have several causes. Some of them are related to the unavailability of radiology specialists, the increasingly overwhelming number of Chest X-Ray images generated daily, or the respiratory disease by itself. Consequently, it might lead to radiologist burnout. Furthermore, an unusual location of pathological lung signs may lead to a misdiagnosis of severe pulmonary disease. Several pathological signs are observed from a Chest X-Ray and on different tissues: Heart and Pericardium, Hili, Mediastinum, Lungs, Pleura, and Chest wall. In the present study, a Deep Learning model is presented, which identifies several tissues and abnormal signs that may require a deeper evaluation. The model was trained using a total of 240,000 Chest X-Ray images, which can automatically identify several structures observed on a Chest X-Ray and some pathological signs of respiratory disease. The model reached 97% of accuracy. These results suggest that our model can be used as an automatic identifier of tissues and pathological signs from Chest X-Ray images. In consequence, it constitutes a meaningful and essential tool for preliminary studies of the severe respiratory disease since it can potentially decrease the number of false positives in the most common radiological examination ordered for patient evaluation. It also could be used for screening purposes to detect findings that may be useful in any pathology assessment, and in the decision-making process of conducting further image modality evaluations, such as CT.