A Deep Network Incorporating Depthwise Separable Convolutions for Pathological Diagnosis of Chest X-Ray Images: A Model Development and Validation Study.

Background: Pathological diagnosis of chest X-ray images has always been a very challenging subject. Methods: We propose a chest X-ray pathology detection network that fuses two depthwise separable convolutions (TDCheXNet) to detect pathology from chest X-ray images. We remove the transformation layer used by the original CheXNet and embed it in depthwise separable convolutions for down-sampling. In the first layer of convolution, to extract more pathological information, we use another depth-separable convolution to replace it and make full use of the negative X-axis features in the corresponding convolution layer. Results: The ChestX-ray14 public chest X-ray dataset is used, which contains more than 100,000 frontal X-ray images covering 14 diseases. Using the area under the receiver operating characteristic curve (AUROC) as the evaluation index, calculate the average AUROC (AVG_AUROC). Experimental results show that TDCheXNet achieved 82.8% on the test set and the detection speed reached 178.794 ms, compared with the original model, AVG_AUROC is improved by 0.5%, and the single-image inference speed is increased by 14.946 ms. Conclusion: We propose a new network for chest X-ray pathology detection, and experimental results show that it can achieve better performance on public datasets.