EBHI: A new Enteroscope Biopsy Histopathological H&E Image Dataset for image classification evaluation

Area of Science:

• Oncology
• Digital Pathology
• Medical Imaging

Background:

• Colorectal cancer is a leading global cancer, necessitating early detection.
• Histopathological examination is crucial for colorectal cancer screening.
• Limited availability of colorectal cancer histopathology datasets, particularly from enteroscope biopsies, impedes AI development.

Purpose of the Study:

• To introduce a novel, publicly available dataset of colorectal cancer histopathology images.
• To facilitate the development and evaluation of computer-aided diagnosis (CADx) techniques for colorectal cancer.
• To address the need for robust datasets for training and testing medical image classification algorithms.

Main Methods:

• A new Enteroscope Biopsy Histopathological H&E Image Dataset (EBHI) was compiled and released.
• The dataset comprises 5532 images across four magnifications and five tumor differentiation stages.
• Machine learning, convolutional neural networks, and transformer-based classifiers were employed for evaluation at 200× magnification.

Main Results:

• Deep learning methods demonstrated superior performance on the EBHI dataset.
• Classical machine learning achieved a maximum accuracy of 76.02%.
• Deep learning models reached a maximum classification accuracy of 95.37%.

Conclusions:

• The EBHI dataset is the first publicly available resource for colorectal histopathology enteroscope biopsies.
• This dataset can spur research into advanced classification algorithms for automated colorectal cancer diagnosis.
• The EBHI dataset has the potential to significantly aid physicians and patients in clinical settings.