Analysis And Recognition Of Pneumoconiosis In X-ray Images Based On Transfer Learning

Pneumoconiosis is an occupational disease with the highest incidence rate,and its diagnosis mainly relies on the analysis of chest X-ray films.However,the mass data of different quality brings heavy stress to the doctors,which is easy to cause misdiagnosis.As pneumoconiosis is irreversible,early screening is particularly important.In modern medical treatment,the use of computer-aided medical image analysis can speed up the diagnosis.But there are still problems of low image quality and low data volume,resulting in a low recognition rate.Therefore,the study of pneumoconiosis identification of X-ray films is carried out around these problems.To solve the problem of the low quality of X-ray films,an improved method of adding local mean-variance to the traditional mean filter is used to remove image noise,and histogram equalization method combined with adaptive gamma correction technology is used to enhance image contrast.These image processing methods can improve the image quality and facilitate the subsequent recognition of pneumoconiosis.Then,the data volume is amplified in a variety of ways to increase sample diversity.To solve the problem of pneumoconiosis data shortage,a transfer learning method including “frozen layers” and “finetuned layers” transfer patterns is proposed to identify pneumoconiosis.The two patterns are built on the pre-trained convolutional neural network,and the transfer is realized by the different updated forms of parameters.The “frozen layers” pattern only updates part of the network layer,while the “finetuned layers” pattern updates the early and subsequent layers of the network with different learning rates.This transfer learning method improves the recognition rate of pneumoconiosis through knowledge in other fields.The X-ray film dataset is used for training and testing.The experimental results show that the accuracy of the data processed by the proposed scheme is significantly improved compared with the original data,especially the method of data amplification,which improved the accuracy from 73.68% to 83.55%.Experimental results also demonstrate that the recognition method based on transfer learning is better than the zero-based method in multiple pre-trained networks.Especially in the Inception-V3 network,the “finetuned layers” transfer pattern achieves 0.94 AUC,higher than o.78 of the method from scratch.