An ensemble WideResNet learning-based approach for classification of multi-class colorectal cancer tissue types in histology images

Abstract

Histopathology imaging (HI) plays a significant role in enhancing the prognosis of colorectal cancer (CRC), which ranks as the second leading cause of cancer-related deaths globally. Classifying colon cancer tissues with HI can be challenging due to differences in morphology, the presence of artifacts when recording microscopic images, and the lack of histological expertise. The use of WideResNet network structure as a novel method for identifying textures in HIs is proposed using deep feature maps extracted from direct HIs and correlation matrices inputs. By using HSV (Hue, Saturation and Value) space, the first step is to decrease various artifacts during HI recording. After designing the ensemble model, the suggested structure has been evaluated and validated using the datasets CRC-5000 and NCT-CRC-HE-100K. On the CRC data set, the ensemble model had accuracy of 98.71% and 99.13%. Deep ensemble learning performed better than current methods in terms of computational and temporal performance, according to the results. Our proposed method is generalizable and has been tested on many of unseen HIs. Based on their data, pathologists can classify unseen Whole Slide Images (WSIs) images using only their lesion classes. The proposed tissue analysis mechanism reliably predicts CRC when pathological images are accurate.