Machine learning-based endoscopic classification for superficial mucosal lesions in hereditary diffuse gastric cancer

Aims Hereditary diffuse gastric cancer (HDGC) is typically due to pathogenetic variants in CDH1 (CDH1-PV) and carries 40-60% life-time risk of signet ring cell carcinoma (SRCC). Endoscopic surveillance can be elected to inform the time of prophylactic total gastrectomy by detecting early SRCC. The pale area is the most common endoscopic lesion harbouring SRCC but can be difficult to diagnose. We previously described the Cambridge criteria to aid endoscopic detection and characterisation of early SRCC in pale areas. [1] The criteria include six features: A, round shape (opposed to linear); B, well demarcated borders (opposed to faded margins); C, focally irregular microvessels; C++, diffusely irregular microvessels; D, irregular microsurface pattern; and E, reproducibility in dynamic view. The aim of this study was to validate the Cambridge criteria prospectively and to develop machine-learning (ML) tool based on these criteria to improve the effectiveness of the endoscopic diagnosis.

Methods CDH1-PV carriers undergoing endoscopic surveillance were prospectively recruited at a single institution between Jan 2020 and Aug 2023. Endoscopies were performed with white light and narrow band imaging with magnification. Pale areas were labelled by three HDGC experts during live examination or post-hoc video analysis using endoscopic features and blinded to current endoscopy histology. The performance of the criteria was analysed based on the sum number of positive features (simple score), individually or as a panel. Afterwards the cohort was randomly split into training and test sets and an explainable ML model, decision tree (DT), was trained and tested based on experts’ labels. Model training and testing were repeated 100x on randomly split datasets, and the performance was evaluated with mean and 95% Confidence Interval (CI). By analysing the ML's diagnostic logic and integrating it with experts’ experience, we generated a clinically applicable DT rule.

Results Overall, 79 CDH1+individuals (60.8% females) were enrolled with an average age of 41.5yr. In total, 215 endoscopic lesions (pale areas) were included from 132 endoscopies, of which 54 (25.1%) were pathologically confirmed SRCC. The performance of each feature is shown in Table 1. Notably all neoplastic lesions were reproducible from different visualisation angles upon dynamic imaging (E+). Using a threshold of≥3 positive features, the simple score achieved an accuracy of 81.4%, sensitivity of 85.2% and specificity of 80.1% in the cohort. The ML model achieved a significantly higher accuracy (84.4%, 95% CI: 83.7%-85.1%) than the simple score (81.7%, 95% CI: 81.1%-82.4%, p<0.001) in the test sets. The feature importance ranked by the decision tree is C>B>A>C++>D. An easy-to-use DT was derived from the diagnostic logic of ML.

Conclusions Specific features of pale areas can be combined to give high accuracy in predicting SRCC in HDGC. A user-friendly DT based diagnostic rule potentially improves performance of the criteria. A prospective validation of the DT is ongoing in an independent cohort.

Conflicts of interest

Authors do not have any conflict of interest to disclose.

• References

• 1 Lee CY, Olivier A, Honing J. et al. Endoscopic surveillance with systematic random biopsy for the early diagnosis of hereditary diffuse gastric cancer: a prospective 16-year longitudinal cohort study. The Lancet Oncology 2023; 24: 107-16

Publikationsverlauf

Artikel online veröffentlicht:
15. April 2024