Advancing Automatic Gastritis Diagnosis: An Interpretable Multilabel Deep Learning Framework for the Simultaneous...

Area of Science:

• Digital pathology
• Artificial intelligence in medicine
• Gastritis diagnosis

Background:

• Accurate diagnosis of gastritis relies on evaluating morphologic features like inflammation, gastric atrophy, and intestinal metaplasia.
• Current artificial intelligence (AI) applications for nontumor diseases, including gastritis, are limited.
• Existing deep learning models often overlook key morphologic indicators and lack simultaneous diagnosis capabilities or interpretable outputs.

Purpose of the Study:

• To develop an attention-based multi-instance multilabel learning network (AMMNet) for simultaneous, interpretable diagnosis of gastritis indicators.
• To evaluate AMMNet's performance in real-world settings by assessing its impact on junior pathologists' diagnostic accuracy and efficiency.
• To validate AMMNet's effectiveness across multicenter data sets.

Main Methods:

• Development of AMMNet, an attention-based multi-instance multilabel learning network.
• Utilized slide-level weak labels for training the model to diagnose activity, atrophy, and intestinal metaplasia.
• Designed a diagnostic test involving junior pathologists with and without AMMNet assistance.

Main Results:

• AMMNet demonstrated high performance in assessing activity (AUC, 0.93), atrophy (AUC, 0.97), and intestinal metaplasia (AUC, 0.93).
• Low false-negative rates were observed: 0.04 for activity, 0.08 for atrophy, and 0.18 for intestinal metaplasia.
• AMMNet assistance reduced junior pathologists' false-negative rates and decreased processing time per whole slide image from 5.46 to 2.85 minutes.
• Block-level clustering analysis provided interpretable visualizations of relevant image regions.

Conclusions:

• AMMNet effectively and accurately evaluates key morphologic indicators for gastritis diagnosis using multicenter data.
• The AI model significantly enhances diagnostic accuracy and efficiency for pathologists, particularly junior ones.
• Multi-instance multilabel learning strategies show promise for supporting routine diagnostic pathology and warrant further investigation.