Image Segmentation for Colorectal cancer histopathological images analysis

Abstract

Colorectal cancer (CRC) is the third most common malignancy and the second most deadly cancer. The most efficient way to determine CRC staging is to analyze whole slide digital pathology images; therefore, it is certainly important to ensure the accuracy of pathology slide analysis.We can obtain medical quantized data of pathological images by implementing deep learning methods. These methods not only can light pathologists’ load but also can provide accurate computing results.In this paper, we use U-2-NET as our backbone to perform Binary Image Segmentation on CRC pathology slides. CRC pathology slides have a variety of non-conforming shapes and colors which is an enormous challenge for detecting cancer areas. U-2-NET was originally used in the Salient Object Detection (SOD) task to find the most unique regions of human attention, which can be used to identify abnormal regions in pathological slices. Moreover, the RSU block of U-2-NET can handle long-term and short-term dependencies, which we believe helps maintain contextual information. With the large computational costs, U-2-NET is hard to implement for application. Our purposed method can use preprocessing, image-selecting mechanisms and transfer learning concepts to solve this problem.Our results show that the model trained with a small part of the data set and a modified small object function has the best results for Binary Image Segmentation of colorectal cancer pathology sections by U-2-NET, with the best IOU (0.77) and Dice Loss (0.83) compared with other models (MSRFCNN, FCN, SegNet, and Unet). Furthermore, after transferring learning using pre-trained weights from the SOD dataset, the results are improved compared to those of learning the network from scratch.