Whole-Case Artificial Intelligence for Diagnostic Escalation in Endoscopically Suspected Borrmann Type IV Gastric Cancer

Technical Summary: Whole-Case Artificial Intelligence for Diagnostic Escalation in Endoscopically Suspected Borrmann Type IV Gastric Cancer

Problem Statement
Borrmann type IV advanced gastric cancer (B-IV AGC), also known as linitis plastica or scirrhous gastric cancer, presents a significant diagnostic challenge. Unlike focal lesions, B-IV AGC is a diffusely infiltrative malignancy that mimics hypertrophic gastritis on endoscopy, characterized by diffuse fold thickening, poor distensibility, and irregular folds. Conventional forceps biopsies frequently yield false-negative results (approximately 50% in non-ulcerated cases) because tumor infiltration is predominantly submucosal. This leads to delayed histologic confirmation, often resulting in patients presenting later with metastatic disease where curative resection is no longer feasible.

Existing artificial intelligence (AI) approaches in gastrointestinal endoscopy have largely relied on curated single-image datasets of focal lesions. This methodology is ill-suited for B-IV AGC, which is defined by examination-level features rather than discrete lesion-level features. Furthermore, prior AI studies specific to B-IV AGC have often used normal or unrelated benign controls and selected representative images, failing to address the clinically critical task of distinguishing true B-IV AGC from suspicious benign mimickers at the whole-case patient level.

Methodology
The authors developed a retrospective, two-center diagnostic modeling study involving 4,860 patients across three cohorts:

1. Primary Cohort: 1,137 patients with endoscopically suspected B-IV AGC (1,137 confirmed B-IV AGC; 319 benign mimickers).
2. Screening Cohort: 3,325 patients (292 B-IV AGC; 3,033 benign gastritis) representing a broader benign comparator.
3. External Validation Cohort: 79 patients from an independent center (47 B-IV AGC; 32 controls).

Model Architecture:
The proposed solution is a "whole-case" patient-level AI model that interprets every available endoscopic image from an examination without manual frame selection or representative image curation.

• Foundation Model: The system utilizes GastroFM, an endoscopy foundation model developed by the authors, to extract image-level feature vectors.
• Aggregation Mechanism: An Attention-Based Multiple Instance Learning (ABMIL) module aggregates these image-level features. The ABMIL assigns learnable attention weights to each image within a patient case, performing a softmax-normalized weighted sum to generate a single patient-level representation.
• Output: A final output layer produces a patient-level probability of B-IV AGC.
• Training Strategy: Cases were split 70:15:15 (train/validation/test). During training, cases with >16 images were randomly sampled to 16 to fit memory constraints, while validation and testing utilized all available images. Class imbalance was addressed via inverse class-frequency weighted random sampling.

Evaluation Protocol:

• Reader Study: Three expert endoscopists (≥10 years experience, >5,000 exams) independently reviewed the held-out primary test set blinded to the reference standard. A "majority-reader" decision was defined as positive if at least two of three experts labeled the case positive.
• Metrics: Performance was assessed using AUROC, sensitivity, specificity, accuracy, PPV, NPV, and decision-curve analysis. A prespecified probability threshold of 0.5 was used for all comparisons.
• Interpretability: Spatial attention heatmaps were generated to visualize which image regions contributed to the prediction.

Key Results

• Primary Cohort Performance: At the 0.5 threshold, the model achieved an AUROC of 0.965 (95% CI 0.936–0.986), with 97.0% sensitivity and 79.6% specificity.
• Comparison with Experts:
  • The model matched the performance of the majority-reader decision (p = 0.143) and significantly outperformed each individual expert (p < 0.05).
  • Clinical Yield: The model identified 163 of 168 pathology-confirmed B-IV AGC cases. This is 5 more cases than the majority-reader decision (158) and 8 to 24 more cases than any individual expert (who detected 144–155 cases).
• Screening Cohort: The model demonstrated even stronger discrimination with an AUROC of 0.998, achieving 95.5% sensitivity and 99.1% specificity.
• External Validation: The model achieved an AUROC of 1.000 on the independent external cohort (47 B-IV AGC, 32 controls), correctly classifying all B-IV AGC cases. However, the authors note the modest sample size (n=79) limits claims of saturated performance.
• Error Analysis:
  • False Negatives: In the 5 missed cases, attention was broadly distributed without focal hyper-attentive frames, consistent with biologically diffuse infiltration lacking discriminative cues.
  • False Positives: In the 10 primary false positives, the dominant pattern was extensive antral hypertrophic fold change, a morphology known to mislead human experts.
  • Anatomy Issue: One external false positive was identified as an esophageal view (superficial esophageal cancer), highlighting that the current model is not anatomy-specific and requires preprocessing to restrict inference to gastric frames.

Significance and Claims
The paper claims that this whole-case AI model addresses a specific "clinical blind spot": the interval between visual suspicion of B-IV AGC and histologic confirmation. By operating on the entire examination rather than curated frames, the model mimics the endoscopist's holistic interpretation process.

The primary significance lies in diagnostic escalation. The model's ability to detect additional pathology-confirmed cases beyond expert consensus suggests it could prompt earlier repeat biopsies, bite-on-bite sampling, or endoscopic ultrasound-guided tissue acquisition in patients with non-diagnostic initial biopsies but high-suspicion endoscopic features. This could reduce the time to histologic confirmation and potentially increase the rate of curative-intent gastrectomy.

The authors emphasize that the model is designed as a decision-support aid for the post-suspicion workflow, not a replacement for histology. They conclude that prospective evaluation is warranted to validate the model's impact on reducing missed and delayed diagnoses in routine practice, noting that anatomy-aware preprocessing and prospective recalibration are prerequisites for clinical deployment.