A Prospective Validation and Observer Performance Study of a Deep Learning Algorithm for Pathologic Diagnosis of...

Area of Science:

• Gastroenterology
• Oncology
• Pathology
• Artificial Intelligence
• Medical Imaging

Background:

• Gastric cancer is a leading cause of death in Northeast Asia.
• Endoscopic screenings are effective for early gastric tumor detection.
• Increasing screening rates necessitate automated diagnostic systems to manage workload.

Purpose of the Study:

• Develop and assess an automated algorithm for classifying gastric epithelial tumors in biopsy images.
• Evaluate the algorithm's performance on a large dataset of gastric biopsies.
• Determine the benefits of the algorithm as an assistive diagnostic tool.

Main Methods:

• A convolutional neural network algorithm was developed using 2,434 whole-slide images.
• The algorithm classified images into three categories: negative for dysplasia (NFD), tubular adenoma, or carcinoma.
• Performance was validated on 7,440 prospective biopsy specimens and assessed by pathologists on 150 cases.

Main Results:

• The algorithm achieved an AUROC of 0.9790 for NFD versus positive classification.
• For epithelial tumors, sensitivity was 1.000 and specificity was 0.9749.
• Algorithm-assisted diagnosis reduced review time by 47% (digital viewer) and 58% (microscopy).

Conclusions:

• The algorithm demonstrates high accuracy in classifying gastric epithelial tumors.
• It functions effectively as an assistance tool for pathologists.
• This system has the potential to serve as a screening aid for gastric biopsy diagnosis.