Computer-Aided Diagnosis in Chest Radiography with Deep Multi-Instance Learning

The Computer-Aided Diagnosis (CAD) for chest X-ray image has been investigated for many years. However, it has not been widely used since limited accuracy. Deep learning opens a new era for image recognition and classification. We propose a novel framework called Deep Multi-Instance Learning (DMIL) on chest radiographic images diagnosis, which combines deep learning and multi-instance learning. Besides, we preprocess images with the alignment based on the key points. This framework can effectively improve the diagnosis effect in the image level annotation. We quantify the framework on three datasets, respectively with different amounts and different classification tasks. The proposed framework obtained the AUC of 0.986, 0.873, 0.824 respectively in classification tasks of the enlarged heart, the pulmonary nodule, and the abnormal. The experiments we implement demonstrate that the proposed framework outperforms the other methods in various evaluation criteria.