medRxiv preprint available

GutCore: An Endoscopy Foundation Model for Whole-Case Gastric Cancer Analysis

Sehun Kim, Hyosoon Yoo, Seong-Keun Yoo, Jeeyun Lee, Yang Won Min, Hyuk Lee
Samsung Medical Center

Study in brief

GutCore was pretrained on 5.6 million de-identified endoscopic images. The study tested whether all images from one endoscopic examination could be aggregated for cancer detection, invasion-depth assessment, biomarker-associated phenotypes, and survival stratification.

Public image-level benchmarks were used as supporting evaluations. GutCore remains a retrospective research model and is not intended for clinical decision-making without further validation.

Select frames, replay the output

Click or drag de-identified frames to replay case-level probabilities and attention maps.

Image bank

Attention analysis

Attention

Selected endoscopy frame
Original frame
Spatial attention heatmap
Spatial attention heatmap
Frame
Image score

Video attention replay

Spatial attention heatmaps

Heatmaps are generated from cancer-detection model attention scores over the same endoscopy clip.

Whole-case clinical evaluation

Diagnosis

Cancer detection

Non-cancer vs pathology-confirmed gastric cancer.

AUC 0.995

Depth

Muscularis propria invasion

Early vs advanced gastric cancer.

AUC 0.960

Biomarkers

EBV and MLH1-associated phenotypes

Tissue-defined biomarker status from whole-case images.

AUC 0.831 / 0.854

Prognosis

Advanced gastric cancer survival

Risk scores for overall survival in AGC.

C-index 0.681

Citation

@article{kim2026gutcore,
  title   = {GutCore: An Endoscopy Foundation Model for Whole-Case Gastric Cancer Analysis},
  author  = {Kim, Sehun and Yoo, Hyosoon and Yoo, Seong-Keun and Lee, Jeeyun and Min, Yang Won and Lee, Hyuk},
  journal = {medRxiv},
  year    = {2026},
  doi     = {10.64898/2026.07.01.26356993},
  url     = {https://www.medrxiv.org/content/10.64898/2026.07.01.26356993v1}
}