Automated Chest Radiographs Triage Reading by a Deep Learning Referee Network

Chest radiographs are often obtained as a screening for early diagnosis tool to rule out abnormalities mainly related to different cardiovascular and respiratory diseases. Reading and reporting numerous chest radiographs is a complex and time-consuming task. This research proposes and evaluates a deep learning (DL) approach based on convolutional neural networks (CNN) combined with a referee fully connected neural network as a computer-aided diagnosis tool in chest X-ray triage and worklist prioritization. The CNN models were trained with a combination of three large scale databases: ChestX-ray14, CheXpert and PadChest. The final database contained 327,176 images labeled with findings obtained by natural language processing (NLP) techniques applied to the radiology reports. The dataset was split in 16 different balanced binary partitions, which were used to train 16 finding-specific classification CNNs. Afterwards, a normal vs abnormal partition of the dataset was created, being abnormal the presence of at least one pathologic change. This final partition was used to train a fully connected neural network as referee that was fed with all the 16 previously trained outcomes. The Area Under the Curve (AUC) analysis evaluated and compared the performance of the models. The system was successfully implemented and evaluated with a test set of 3400 images. The AUC of the normal vs abnormal classification was 0.94. The highest AUC of the finding-specific classifiers was 0.99 for hernia. The proposed system can be used to assist radiologists identifying abnormal exams, allowing a time-efficiency triage approach.