Weakly Supervised Learning for Poorly Differentiated Adenocarcinoma Classification in GastricEndoscopic Submucosal...

Area of Science:

• Computational pathology
• Digital pathology
• Oncology

Background:

• Endoscopic submucosal dissection (ESD) is key for early gastric cancers, particularly poorly differentiated adenocarcinoma.
• Accurate histopathological classification is crucial for treatment and patient outcomes.
• Manual microscopic analysis of whole-slide images (WSIs) is time-consuming and resource-intensive.

Purpose of the Study:

• To develop a computer-aided diagnostic system for classifying gastric poorly differentiated adenocarcinoma from ESD WSIs.
• To leverage deep learning for rapid and accurate analysis of histopathological specimens.
• To assist pathologists in routine diagnostic workflows.

Main Methods:

• Training a deep learning model using transfer and weakly supervised learning approaches.
• Classifying poorly differentiated adenocarcinoma in ESD WSIs.
• Utilizing computational pathology techniques for image analysis.

Main Results:

• The deep learning model achieved a high ROC-AUC of up to 0.975 on gastric ESD WSI test sets.
• The model demonstrated strong performance across ESD, endoscopic biopsy, and surgical specimen WSI datasets.
• Successful classification of poorly differentiated adenocarcinoma was achieved.

Conclusions:

• The developed deep learning model shows significant potential for integration into practical gastric ESD histopathological diagnostic workflows.
• This computational pathology system can serve as a valuable computer-aided diagnosis tool.
• The study highlights the benefits of AI in accelerating and improving cancer diagnosis.