Automated detection of lung lesions on Chest X-ray images shows good performance to\nreduce lung cancer mortality. However, it is difficult to detect multiple lesions of single image well\nand truly, and additional efforts are needed to improve diagnostic efficiency and quality. In this paper,\na multi-label classification model combining attention-based neural networks and association-specific\ncontexts is proposed for the detection of multiple lesions on chest X-ray images. A convolutional\nneural network and a long short-term memory network are first aligned by an attention mechanism to\ntake advantage of both image and text information for the detection, called CNN-ATTENTION-LSTM\n(CAL) network. In addition, a mining method of implicit association strength to obtain an association\nnetwork of chest lesions (CLA) network is designed to guide the training of CAL network. The CLA\nnetwork provides possible clinical relationships between lesions to help the CAL network obtain\nbetter predictions. Experimental results on ChestX-ray14 dataset show that our method outperforms\nsome state-of-the-art models under the metrics of area under curve (AUC), precision, recall, and\nF-score and achieves up to 85.4% in the case of atelectasis and infiltration. It indicates that the method\nmay be useful in the computer-aided detection of multiple lesions on chest X-ray images.
Loading....