Research On Deep Convolutional Neural Networks Based Pathological Image Diagnosis

Pathological image diagnosis is a key problem in medical image analysis.Due to the subjective factors of pathologists,manual diagnosis is time consuming and has limited accuracy.Computer-aided methods have been developed to alleviate pathologists’ burden.Diseased and healthy tissues show distinct difference in morphology,so extracting morphological features of cells is important to accurate pathological diagnosis.This thesis leverages Convolutional Neural Network(CNN)as a tool for feature extraction and presents two CNN based models for automatic pathological image diagnosis.One of the models combines CNN with ensemble learning,and the other combines CNN with multi-scale image analysis.The thesis first reviews the literature on pathological images diagnosis.Then,we describe the theories and application scenarios of CNN,ensemble learning and multiscale image analysis.We propose a boosting convolutional neural network based algorithm for pathological image diagnosis.The algorithm uses two heterogeneous networks,that are the basic classifier and the boosting network,to learn abstract features from pathological images.The basic classifier aims to predict the probabilities of cells being tumor,and the boosting network corrects the predictions made by the basic one.To reduce the risk of over-fitting caused by scarce of training examples,we augment training samples and impose a consistency constraint on training objective,forcing the classification results of augmented samples consistent with those of the original ones.Based on CNN and multi-scale image analysis,the thesis also presents a multiscale squeeze-and-excitation network based pathological image diagnosis algorithm.The proposed algorithm uses the densely connected convolutional network with squeeze-and-excitation modules to learn the features of pathological images at a certain scale.The squeeze-and-excitation module adaptively gives more emphases to the features that can lead to more accurate diagnosis.The feature learning network was applied on pathological images in parallel to obtain their multi-scale representations.The efficacies of the two algorithms were tested on the public databases of pathological images of mammalian organs and lymph nodes of breast cancer.To tackle the complicated structures of lymph nodes in breast cancer tissues,we design a postprocessing scheme for image-level diagnosis.We aggregate the probabilities that local patches being tumor tissues in a pathological image to a heatmap,and then extract a set of geometrical features from the connected components in heatmap.Subsequently,a boosting tree is trained using the geometrical features extracted from heatmap to estimate the probability that an image contains cancer tissues.Finally we evaluate diagnosis accuracy on image-level.Experimental results indicate that the accuracies of two proposed models are both higher than 90%.