Histopathological Colorectal Cancer Classification And Survival Analysis With Self-supervised Learning

Deep learning has made great success in histopathological images classification.Using deep learning methods to detect and classify dieases in histopathological images is of great significance.However,the performance of deep learning approaches rely heavily on the substantial task-specific annotations,which is affected by the quality of the histopathological images,and the data annotation require experienced pathologists’ manual labelling.As a result,they are laborious and time-consuming.However,unlabeled histopathological data is easy to obtain.Therefore,how to capitalize on the abundance of unlabeled data to improve the performance of a histopathological task is still an challenging problem.To solve this problem,this study explores the histopathological image classification and survival analysis model of colorectal cancer based on self-supervised transfer learning.The specific work content is as follows:(1)We propose a self-supervised Deep Adaptive Regularized Clustering(DARC)framework to pre-train a neural network.For the colorectal cancer classification task,we frist pre-train a neural network and adopt it as initial weights to train the histopathological image classification model.For the patient stratification task,the proposed DARC is used to obtain the representations of histopathological images,the clusters after clustering the representations of histopathological images are evaluated on survival in combination with clinical data to explore the impact of the learned clusters on survival.DARC iteratively clusters the learned representations and utilizes the cluster assignments as pseudo-labels to learn the parameters of the network.To further improve the model performance,we propose a clustering loss that measures the distance between data representations and the corresponding clusters to make the data representations in the same cluster be close as much as possible.In addition,we design an objective function combining a network loss with a clustering loss using an adaptive regularization function,which is updated adaptively throughout the training process to learn feasible representations.For the colorectal cancer classification task,the proposed DARC is evaluated on dataset of NCT-CRC-HE-100 K.The accuracy of using the network trained using DARC pre-trained weights with only 10% labeled data(98.57%)is already comparable to the network trained from scratch with 100% training data(98.08%).For the patient stratification task,histomorphological clusters obtained by DARC on the slice data and clinical data of 454 patients with colorectal cancer are evaluated by training survival model,the global concordance index(cindex)of the survival analysis model is 0.73.(2)We propose a self-supervised learning method based on Unsupervised Adversarial Constrastive Learing(UACL)to pre-train a neural network,and adopt the pre-trained weights to histopathological image classification task and patient stratification task.UACL uses adversarial training to generate more challenging and difficult adversarial sample pairs in the representation space,so that to learn more robust representations.For the colorectal cancer classification task,the proposed DARC is evaluated on dataset of NCT-CRC-HE-100 K.The results show that the proposed UACL with only 1% is already comparable to the network trained from scratch with 100% training data.For the patient stratification task,histomorphological clusters obtained by DARC on the slice data and clinical data of 454 patients with colorectal cancer are evaluated by training survival model,the global concordance index(c-index)of the survival analysis model is 0.75.The experimental results show the survival model have statistically significant patient stratification.