Supervised Constructive Learning-Based Model for Identifying Colorectal Cancer Tissue Types From Histopathological Images

Abstract

Colorectal cancer is the disease with the second highest mortality rate among cancer types. The survival rate is increased with early diagnosis and treatment of this disease. In this study, a supervised constructive learning based model is proposed for the detection of colorectal cancer using datasets containing hematoxylin and eosin stained colon histopathological images. The datasets used include multi-class datasets (Kather-5K, CRC-7K, NCT-100K) and binary class datasets (Kather MSI and MHIST). The proposed model consists of an encoder (ReFeatureBlock (RFB), depthwise convolution (DWC), and global average pooling (GAP)), a projection head, and fully connected classification networks. With these networks, it is possible to obtain important features, reduce the computational cost, minimize noise sensitivity, and prevent poor margin possibilities. Additionally, the Grad-CAM method was used to ensure transparency and explainability of the model's decision-making processes. In multiple classification experiments, in applications performed by combining Kather-5K, CRC-7K, and NCT-100K datasets, the proposed model achieved the highest performance with 99.21% accuracy, 99.19% precision, 99.19% recall, 99.19% F1-score, 99.92% specificity, and 99.56% AUC values, respectively. In addition, in tests performed on individual datasets, high performances such as 99.10% accuracy for Kather-5K, 99.76% accuracy for CRC-7K, and 99.19% accuracy for NCT-100K were achieved. In binary classification experiments with the MHIST dataset, the proposed model showed the highest success with 99.52% accuracy, 99.30% precision, 99.49% recall, 99.40% F1-score, 99.49% specificity, and 99.49% AUC, respectively. Moreover, the proposed model is compared with state-of-the-art techniques in the literature in the classification of colorectal cancer tissues, and the results are discussed. The findings show that the proposed model provides higher classification success in statistical metrics. The codes of the proposed model are publicly available at .