A foundation-model-powered AI system with post-read alerts and concurrent assistance improves gastric biopsy diagnosis across 29 pathological categories

Technical Summary: Foundation-Model-Powered AI for Gastric Biopsy Diagnosis

Problem Statement
Gastric biopsy interpretation is a complex diagnostic task requiring the simultaneous assessment of malignancies, precancerous lesions (dysplasia, metaplasia), infections (Helicobacter pylori), and various benign conditions. Current artificial intelligence (AI) models in this domain are limited by narrow scope, typically addressing only two to three diagnostic categories (e.g., malignant vs. non-malignant), and have been evaluated under a single mode of clinical integration. Furthermore, traditional deep learning approaches for multi-class classification require extensive pixel-level annotations or massive datasets, creating barriers to scalability and deployment in diverse diagnostic settings. There is a lack of empirical evidence comparing different workflow integration paradigms (e.g., concurrent assistance vs. post-read alerts) within a single framework for gastric pathology.

Methodology
The study developed a deep learning system leveraging a Vision Foundation Model (VFM) to address these limitations.

• Model Architecture and Pre-training: The system utilizes a ResNet-50 architecture pre-trained on a large-scale, diverse dataset of 70,618 whole slide images (WSIs) from multiple organ systems and disease types using Barlow Twins contrastive learning. This pre-training was performed on a self-supervised basis without task-specific labels.
• Fine-tuning Strategy: The pre-trained VFM was fine-tuned end-to-end for a 29-class multi-label classification task using the Enhanced Streaming CNN (ESCNN) method. This approach allows for whole-slide-level training without requiring tile-level or pixel-level annotations. The model was trained on 10,462 gastric biopsy WSIs from Linkou Chang Gung Memorial Hospital (2020–2023).
• Classification Scope: The model classifies biopsies into 29 pathological categories, organized hierarchically into high-priority (malignancy subtypes), medium-priority (dysplasia, metaplasia, H. pylori, benign tumors), and low-priority categories (inflammation, polyps, ulcers, etc.).
• Annotation-Free Localization: To generate pixel-resolution heatmaps without pixel-level supervision, the model employs Rectified Log-Sum-Exp (ReLSE) pooling combined with streaming convolution. This produces Class Activation Maps (CAMs) directly from slide-level labels.
• Clinical Integration Paradigms: Two distinct workflows were evaluated:
  1. Post-Read Alert System: A retrospective simulation where the model flags malignancies missed or miscoded in routine Laboratory Information System (LIS) records across four independent hospital cohorts (n=4,679).
  2. Concurrent Assistance System: A prospective reader study involving six pathologists who reviewed 120 WSIs under two conditions: unaided and with AI assistance (heatmaps for malignancy, dysplasia, metaplasia, and H. pylori).
• Data Efficiency Evaluation: "Few-shot" experiments were conducted to compare the VFM-fine-tuned model against a model trained from scratch using balanced subsets of 200, 400, and 800 slides.

Key Results

• Diagnostic Performance: On an internal test set (n=2,957), the model achieved an Area Under the Curve (AUC) of 0.9945 for malignancy detection. It also demonstrated high performance for dysplasia (AUC 0.9707), metaplasia (AUC 0.9489), and H. pylori (AUC 0.9210).
• Post-Read Alert Efficacy: In the retrospective simulation across four external cohorts, the system identified 17 malignancies that were missed or miscoded in routine diagnostic records. Expert adjudication confirmed these as true positives, demonstrating the system's ability to serve as a quality assurance tool for retrospective error recovery.
• Concurrent Assistance Impact: In the prospective reader study, AI assistance significantly improved pathologist sensitivity for malignancy (increasing from 92.2% to 94.8%, P = .034) and metaplasia (increasing from 94.2% to 97.4%, P = .013). The median review time per slide decreased slightly but significantly from 43.8 seconds to 41.1 seconds (P = .026). Specificity remained statistically unchanged.
• Data Efficiency: A VFM fine-tuned on only 400 slides (200 malignant, 200 benign) achieved a malignancy AUC of 0.9906. This performance surpassed a model trained from scratch on the same 400-slide subset (AUC 0.9660) and also outperformed a scratch-trained model using the full 10,462-slide dataset (AUC 0.9755).
• Localization: The model successfully generated pixel-resolution heatmaps for all 29 categories without any pixel-level annotations, enabling the visualization of co-existing lesions (e.g., malignancy with metaplasia and H. pylori) in a single inference pass.

Significance and Claims
The authors claim this study establishes a framework for dual-mode AI deployment in gastric biopsy pathology, addressing the need for both real-time diagnostic support and retrospective quality assurance. The primary contributions are:

1. Comprehensive Scope: The first gastric biopsy AI model to simultaneously classify 29 pathological categories, mirroring the full scope of pathologist decision-making rather than limiting output to binary malignancy detection.
2. Workflow Validation: The first study to prospectively evaluate both concurrent assistance and post-read alert paradigms within the same framework, demonstrating that these modes offer complementary trade-offs suitable for different clinical settings (high-throughput screening vs. retrospective error correction).
3. Data Efficiency: Systematic quantification showing that VFM pre-training can reduce annotated data requirements by orders of magnitude (e.g., 400 slides vs. 10,000+) while maintaining or exceeding the accuracy of traditional large-scale training.
4. Annotation-Free Localization: Demonstration that pixel-resolution lesion localization across 29 categories is achievable using only slide-level labels, removing the prohibitive cost of pixel-level annotation for scaling AI in pathology.

The authors position these findings as foundational building blocks for future "agentic" AI systems that can autonomously orchestrate pathology workflows, leveraging data-efficient foundations to adapt to rare diseases and diverse histological patterns without extensive re-curation.