Pneumoconiosis identification in chest X-ray films with CNN-based transfer learning

Pneumoconiosis is one of the most universal and severe occupational diseases in the world today. The diagnosis of pneumoconiosis mainly depends on the doctors’ analysis of chest X-ray films. However, diagnostic accuracy is related to a doctor’s expertise. Recent research has suggested that deep convolutional neural network (CNN) could better identify diseases than experts. Deep CNN can efficiently extract features from data for discrimination. Nevertheless, high volumes of training data are usually necessary to achieve, which is not desirable from the research. This article studies pneumoconiosis identification from the perspective of CNN-based transfer learning. In this work, we raise two transfer learning patterns from “frozen layers” and “finetuned layers” to solve those problems. The knowledge learned from the ImageNet is transferred to pneumoconiosis identification. Also, we propose some methods of image denoising, lung segmentation, and data amplification for preprocessing to improve image qualities. The experimental results show that both transfer learning patterns are better than starting from scratch with limited training data. Notably, the “finetuned layers” transfer learning pattern achieves a higher performance.