Transfer Learning For Binary Detection Of Bone Abnormalities On Orthopedic X-ray Images

Musculoskeletal diseases affect more than 1.7 billion people worldwide and are the most common cause of severe long-term pain and disability,with 30 million emergency department visits each year,and this number is increasing.On May 25,2018,Wu Enda and the Stanford ML team opened a dataset of bone medical images MURA.The purpose of this article is to use this data set and use a convolutional neural network to train a binary classification system that can determine whether a patient in this image has an orthopedic disease only by inputting X-ray images.Because the data set itself only gives negative and positive labels(ie,suffering from a disease),this article only discusses in the category of judging yin-positive,and cannot specifically determine which disease it is.The goal of the system is to screen radiologists for negative cases by using a high recall method(correspondingly to reduce accuracy),so that they can focus on cases that the system judges to be positive.The MURA dataset contains 40,000 images,but there is a lot of noise in it.For example: Except for the orthopedic part,the background is more messy;all images are scanned from the printed image,but in this case the image itself already has some letters printed by the radiologist;the contrast of many image key parts The brightness is insufficient.In fact,after the training,I used the heat map to check the convolution kernel's attention.Many times,the background was shared a lot.This thesis mainly completed the following work:1.Increase image contrast.Here I used the adaptive histogram equalization method.2.Segmentation of the image(I want to separate the orthopedic part from the cluttered background).This is the most difficult and time-consuming step in the dissertation.The grab Cut,canny,watershed algorithm,and intermediate generation adversarial network I used from the beginning,but the results show that if there is no human intervention at all,the segmentation situation is far from the expected goal.This is also due to the characteristics of the X-ray image itself: the muscles near the edges are already very dark,which is easily confused with the background.In the end,I used the OPEN operation combinition,and I got good results at some of the pictures.3.Model training: I compared the network structures of different depths in resnet,densenet and Google's latest Efficient Net,and finally found that densenet169 works best.