Advanced deep learning for multi-class colorectal cancer histopathology: integrating transfer learning and ensemble methods

Abstract

Background: Cancer is a major global health threat, constantly endangering people's well-being and lives. The application of deep learning in the diagnosis of colorectal cancer can improve early detection rates, thereby significantly reducing the incidence and mortality of colorectal cancer patients. Our study aims to optimize the performance of deep learning model in the classification of colorectal cancer histopathological images to assist pathologists in improving diagnostic accuracy.

Methods: In this study, we developed ensemble models based on deep convolutional neural networks (CNNs) for the classification of colorectal cancer histopathology images. The method first involved data preprocessing techniques such as patch cropping, stain normalization, data augmentation and data balancing on histopathology images with different magnifications. Subsequently, the CNN models were fine-tuned and pre-trained using transfer learning methods, and models with superior performance were then selected as the base classifiers to build the ensemble models. Finally, the ensemble models were used to predict the final classification outcomes. To evaluate the effectiveness of the proposed models, we tested their performance on a publicly available colorectal cancer dataset, Enteroscope Biopsy Histopathological Hematoxylin and Eosin Image (EBHI) dataset.

Results: Experimental results show that the proposed ensemble model, composed of the top five classifiers, achieved the promising classification accuracy across sub-databases with four different magnification factors. Specifically, on the 40× magnification subset, the highest classification accuracy reached 99.11%; on the 100× magnification subset, it reached 99.36%; on the 200× magnification subset, it was 99.29%; and on the 400× magnification subset, it was 98.96%. Additionally, the proposed ensemble model achieved exceptional results in recall, precision, and F1 score.

Conclusions: The proposed ensemble models obtained good classification performance on the EBHI dataset of histopathological images for colorectal cancer. The findings of this study may contribute to the early detection and accurate classification of colorectal cancer, thereby aiding in more precise diagnostic analysis of colorectal cancer.

Keywords: Histopathological image; colorectal cancer; deep learning (DL); ensemble learning; transfer learning (TL).

Figure 1

Representative EBHI dataset samples. EBHI, Enteroscope Biopsy Histopathological Hematoxylin and Eosin Image; IN, intraepithelial neoplasia.

Figure 2

The workflow of the Patch-Cropping Algorithm processing.

Figure 3

Original and stain normalized image samples.

Figure 4

The architecture of the proposed ensemble model. CNN, convolutional neural network.

Figure 5

Samples and corresponding Grad-CAM heatmaps generated by eight CNN models under the four magnification factor images. Grad-CAM, Gradient Weighted Class Activation Mapping; CNN, convolutional neural network.