Deep-Hipo: Multi-scale receptive field deep learning for histopathological image analysis.

Area of Science:

• Digital pathology
• Computational pathology
• Machine learning in oncology

Background:

• Whole-slide imaging digitization advances computer-aided tissue examination.
• Convolutional neural networks (CNNs) are key for analyzing histopathological images in cancer detection and classification.
• Existing patch-based CNN methods struggle with limited information from small image patches.

Purpose of the Study:

• To introduce Deep-Hipo, a novel multi-task deep learning model for histopathological image analysis.
• To address the limitations of patch-based methods by incorporating multi-scale information.
• To improve the accuracy of cancer detection, risk prediction, and subtype classification in digital pathology.

Main Methods:

• Proposed Deep-Hipo, a multi-task deep learning model utilizing simultaneous multi-scale patches.
• Extracted two patches of the same size at high and low magnification levels from whole-slide images.
• Captured complex morphological patterns using large and small receptive fields.

Main Results:

• Deep-Hipo demonstrated superior performance compared to state-of-the-art deep learning methods.
• The model was successfully assessed on diverse whole-slide gastric adenocarcinoma images (well, moderately, poorly differentiated, and signet-ring cell carcinoma) and normal mucosa.
• Applied to TCGA Stomach Adenocarcinoma (TCGA-STAD) and Colon Adenocarcinoma (TCGA-COAD) datasets, yielding clinically verified results by a pathologist.

Conclusions:

• Deep-Hipo effectively analyzes histopathological images by integrating multi-scale information.
• The model shows significant potential for accurate cancer analysis in digital pathology, outperforming existing methods.
• The source code is publicly available, facilitating further research and application.