Classification For Pathological Images Of Colorectal Cancer With Deep Learning

ObjectiveAs one of the malignant tumors,colorectal cancer seriously endangers human health.So the automatic diagnosis of Hematoxylin-Eosin stained pathological images of colorectal cancer is of great clinical importante.In this study,we aimed to automatically classify the pathological images of normal tissues,tumor tissues,tumor tissues with different differentiation degrees and tumor tissues with different T stages of colorectal cancer using deep learning algorithm and transfer learning,and to explore the accuracy of deep learning algorithm in the classification of colorectal cancer pathological images.MethodsIn this study,pathological full-section scan images of 349 colorectal cancer patients（including 304 moderately differentiated and 40 poorly differentiated patients）from the TCGA database were collected,among which 38 patients had paracancer normal tissue sections.The pathological full-section scan images were used as the input data of the convolutional neural network after dividing into non-overlapping images of 512*512*3 pixel.Finally,1003 normal tissues images and 9187 tumor tissues images were obtained,which were divided into training set,verification set and test set in a ratio of 8:1:1.The Inception V3 transfer learning model was used to build model,and all operations were based on Python’s Tensor Flow framework.Results1.In the Inversion V3 transfer learning model to classify the normal tissues and tumor tissues of colorectal pathology images,the accuracy,sensitivity,specificity,precision,recall,F₁-score,AUC and the area under precision-recall curve on the test set were 99.7%,99.7%,100%,100%,99.7%,0.998,1,and 1 respectively.2.The accuracy,sensitivity,specificity,precision,recall,F₁-score,AUC and the area under precision-recall curve of colorectal cancer pathological images with different degrees of differentiation were 96.4%,97.1%,95.9%,93.9%,97.1%,0.955,0.99 and 0.99 respectively.3.The accuracy of pathological images of colorectal in different T stages was 62.1%in the test set,and the accuracy of T₁-T₄stage were 80.0%,62.2%,61.5%and62.0%,respectively.Conclusion1.The Inception V3 transfer learning model nearly correctly distinguishes normal tissues from tumor tissues in colorectal cancer.The sensitivity,specificity and accuracy achieved are almost consistent with manual diagnosis.2.The Inception V3 transfer learning model correctly distinguishes the majority of moderately differentiated and poorly differentiated colorectal cancer images.3.The diagnostic accuracy of Inception V3 transfer learning model for colorectal cancer with different T stages needs to be improved.