Deep Learning-based Colorectal Pathological Auxiliary Diagnosis

Colon cancer is one of the most common malignant tumors of digestive tract.The incidence and mortality are increasing in China in recent years.The diagnosis depends heavily on the clinical experience of the pathologists.Since doctors have heavy workload,misdiagnosis events occur frequently.Hence,it is urgent to study the computer-aided diagnosis technology for colon cancer to assist pathologists to improve the efficiency of diagnosis and eliminate misdiagnosis.This paper studies the segmentation of glands and the diagnosis of benign and malignant lesions in colorectal histology images.The specific work and results are as follows:First,a method of gland segmentation based on Segnet is proposed.In contrast to the existing Segnet segmentation methods in the literature,the Warwick-QU of colorectal pathological dataset is effectively augmented by cropping.Then,the augmented dataset is used for training Segnet and gland segmentation test.The results show that the segmentation accuracy of 0.882 and shape similarity of 106.6471 are obtained on Part A of Warwick-QU,which is close to the best results of CUM2.The highest segmentation accuracy of 0.8636 and shape similarity of 102.5729 are obtained on Part B,which is higher than that of CUM1 and ExB1 about 0.06 and 43,respectively.Second,an auxiliary diagnosis algorithm based on multi-feature is designed for diagnosis of colon cancer.The SVM diagnosis model based on contour,color and texture features and its combination is constructed and used to diagnose benign and malignant lesions in two data sets: D1(original pathological image dataset)and D2(data set after glandular segmentation).The results show that the pathological diagnosis model has higher diagnostic accuracy on D2.The SVM model based on contour and texture has achieved the highest diagnostic accuracy of 83.75%,which shows that it is hard to use the traditional image processing method to diagnosis the benign and malignant lesions of colon cancer.Finally,a method of colorectal pathological diagnosis based on two kinds of deep convolution neural networks(CIFAR and VGG)is proposed.After configuring and training the two networks,the trained CIFAR and VGG are applied to the diagnosis of D1 and D2.The results show that the diagnostic results of CIFAR and VGG is much higher than that of SVM pathological diagnosis model described by multiple features.The results of CIFAR and VGG are increased by 6.25% and 8.75% on D1 and 10.25% and 12.5% on D2,respectively.VGG has obtained the highest diagnostic accuracy of 96.25% on D2,which is very close to the highest diagnostic accuracy of 97.5% in the existing methods in literature.