Improved classification of colorectal polyps on histopathological images with ensemble learning and stain normalization. (April 2023)

• Record Type: Journal Article
• Authors: Yengec-Tasdemir, Sena Busra; Aydin, Zafer; Akay, Ebru; Dogan, Serkan; Yilmaz, Bulent
• Is Part Of: Computer methods and programs in biomedicine. Volume 232(2023)
• Publication Date: 2023-04
• Subjects: 41A05 -- 41A10 -- 65D05 -- 65D17 -- Colorectal Polyps -- Colonic Polyp Classification -- Histopathology Image Classification -- Computer-aided Diagnosis -- Clinical Decision Support System -- Ensemble of Deep Convolutional Neural Networks -- ConvNeXt -- Transfer Learning; Medicine -- Computer programs -- Periodicals; Biology -- Computer programs -- Periodicals; Computers -- Periodicals; Medicine -- Periodicals; Médecine -- Logiciels -- Périodiques; Biologie -- Logiciels -- Périodiques; Biology -- Computer programs; Medicine -- Computer programs; Periodicals; Electronic journals; 610.28
• DOI: 10.1016/j.cmpb.2023.107441
• Languages: English
• ISSNs: 0169-2607

Abstract

Highlights: Different from the previous studies which make polyp classification on HI, in this study, various stain normalization techniques are combined with an ensemble model. To the best of our knowledge, this study is the first to use ensemble methods to classify colonic histological images as adenomatous and non-adenomatous. In this study, we explored state-of-the-art pre-trained Deep CNN algorithms' performances on our custom dataset. To the best of our knowledge, this study is the first to comprehensively evaluate widely used stain normalization techniques namely, Stain-Gan, Stain-Net, Vahande, Macenko and Reinhard, for classification of adenomatous and non-adenomatous colonic polyp tissues. Moreover, this study is one of the first studies which use ConvNeXt architecture on the histopathologic image classification task. Furthermore, we propose a novel ensemble model which ensembles the pre-trained ConvNeXt-tiny and ConvNeXt-base variants to classify adenomatous and non-adenomatous tissues from colonic histopathology images. In order to comprehensively evaluate and assess the generalizability of the proposed model, during the testing phase, we also employ publicly available UniToPatho and EBHI databases. The proposed ensemble model achieves an accuracy of 95% on our custom dataset. Additionally, in order to ensure the explainability of the proposed model, the Grad-Cam method is used. The attention map of the model is explored for adenomatous and non-adenomatous images. We believe that these Grad-Cam printouts of the proposed model for the pathological images can guide pathologists. Abstract: Background and Objective : Early detection of colon adenomatous polyps is critically important because correct detection of it significantly reduces the potential of developing colon cancers in the future. The key challenge in the detection of adenomatous polyps is differentiating it from its visually similar counterpart, non-adenomatous tissues. Currently, it solely depends on the experience of the pathologist. To assist the pathologists, the objective of this work is to provide a novel non-knowledge-based Clinical Decision Support System (CDSS) for improved detection of adenomatous polyps on colon histopathology images. Methods : The domain shift problem arises when the train and test data are coming from different distributions of diverse settings and unequal color levels. This problem, which can be tackled by stain normalization techniques, restricts the machine learning models to attain higher classification accuracies. In this work, the proposed method integrates stain normalization techniques with ensemble of competitively accurate, scalable and robust variants of CNNs, ConvNexts. The improvement is empirically analyzed for five widely employed stain normalization techniques. The classification performance of the proposed method is evaluated on three datasets comprising more than 10k colon histopathology images. Results : The comprehensive experiments demonstrate that the proposed method outperforms the state-of-the-art deep convolutional neural network based models by attaining 95% classification accuracy on the curated dataset, and 91.1% and 90% on EBHI and UniToPatho public datasets, respectively. Conclusions : These results show that the proposed method can accurately classify colon adenomatous polyps on histopathology images. It retains remarkable performance scores even for different datasets coming from different distributions. This indicates that the model has a notable generalization ability.