Improving histopathological image classification with patch-based embeddings and ensemble learning: a study on enteroscope biopsy images

Abstract Histopathological image analysis is crucial for the accurate diagnosis and classification of cancer types. Traditional methods rely heavily on pathologists’ expertise, which can be subjective and time-consuming. Existing patch-based approaches for WSI classification typically use all patches from an image for training, which is computationally expensive. Previous studies on the EBHI dataset achieved 95.37% accuracy for binary classification and 91.1% for multi-class classification. Current methods often struggle with class imbalance and require extensive computational resources. This study aims to enhance the classification accuracy of Enteroscope Biopsy Histopathological images by employing patch-based embeddings and ensemble learning techniques, focusing on five distinct cancer classes. The research gaps identified are as follows: (1) High computational cost due to processing all patches from WSI images; (2) Limited investigation of optimal patch selection strategies; (3) Insufficient comparison of embedding aggregation methods (average, max, similarity-based); (4) Lack of comprehensive ablation studies examining patch size, feature extractors, and class-weighted penalization effects. The study involves dividing each histopathological image into non-overlapping patches, from which embeddings are extracted using a pre-trained network. These embeddings are then classified using ensemble learning algorithms, including Random Forest, XGBoost, LightGBM, AdaBoost, CatBoost, and Artificial Neural Networks. The effectiveness of the approach is evaluated through extensive experiments, including an ablation study to assess the impact of patch size, feature extractors, and class-based penalization on classification performance. Statistical measures such as precision, sensitivity, and specificity are used to evaluate the models. The proposed method achieved a classification accuracy of 94.5% for multi-class classification and 98.9% for binary-class classification, outperforming existing studies. The ensemble learning approach demonstrated high precision, sensitivity, and specificity, with CatBoost and LightGBM showing superior performance. The study also found that using selected embeddings from patches significantly reduced computational costs while maintaining high classification accuracy. This study demonstrates the potential of patch-based embeddings combined with ensemble learning to improve the classification of histopathological images. The findings suggest that this approach can provide a robust and scalable solution for medical image classification, potentially leading to more accurate and efficient cancer diagnosis.