Dual-Model Transfer Learning with Graph Neural Networks and SHAP Explainability for Robust Colorectal Tumor Margin Detection Process
Journal of Innovative Image Processing · 17.07.2026 · 8 (3) · 1076-1095
DOI: 10.36548/jiip.2026.3.016 ↗Kurz gefasst
Fachartikel in der Zeitschrift Journal of Innovative Image Processing, online erschienen am 17.07.2026. Verfasst von Divya Midhunchakkaravarthy und 2 weiteren Autor:innen. Die Arbeit nennt 19 Literaturangaben.
Abstract (Originalsprache)
When performing surgery for colorectal cancer, it is essential to precisely delineate the margins of the tumor. This is because even minute errors in the margins can have a significant impact on the quality of the resection, the risk of recurrence, and the planning of therapy following the surgery. When slides are supplied from multiple locations, the computer-aided pathology models now in use do not always work correctly, and their explanations do not always focus on the invasive boundary, which is where clinical decisions are most critical. The purpose of this study is to demonstrate a two-model transfer learning system that combines spatial aggregation based on graph neural networks with SHAP-guided border explainability in order to improve the accuracy of colorectal tumor margin identification. Weighted tissue graphs are created from the slides used in histopathology. While the lines illustrate how the cells are related in space shape by utilizing Delaunay and k-nearest neighbor, the nodes illustrate the areas that are split into cellular or super-pixel regions. At the same time, five modules are utilized: CDGCV, which stands for Cross-Domain Graph Consistency Validation, is a technique that verifies the consistency of the topology between centers. BFSRT, which stands for Boundary-Focused SHAP Refinement Test, is a test that verifies the correctness of boundary-specific explanations. The strength of the structure is evaluated using a technique known as Multi-Scale Relational Perturbation Assessment (MSRPA); Dynamic SHAP-Graph Alignment (DSGA) is used to determine whether t the interpretation is timely, and Hierarchical Graph-SHAP Ensemble (HGSE) is used to determine whether the epithelium, stroma, and invasive-front cells are in accord with one another. Dice scores can reach up to 0.94, boundary fidelity is close to 0.90, cross-center consistency is over 0.82, and F1 degradation is less than 3% when edges are rewired 20%, according to experimental tests conducted on the Kather, CRC-TIA, and internal multi-center cohorts. The support for transferability is increased by testing conducted at an external center. Therefore, the system provides accurate margin localization, stable graph reasoning, and explanations that are comprehensible to medical professionals for the purpose of deploying colorectal pathology across many locations.
Zitieren
APA
Midhunchakkaravarthy, D., G., M. N., & V., L. N. (2026). Dual-Model Transfer Learning with Graph Neural Networks and SHAP Explainability for Robust Colorectal Tumor Margin Detection Process. Journal of Innovative Image Processing, 8(3), 1076-1095. https://doi.org/10.36548/jiip.2026.3.016
Vancouver
Midhunchakkaravarthy D, G. MN, V. LN. Dual-Model Transfer Learning with Graph Neural Networks and SHAP Explainability for Robust Colorectal Tumor Margin Detection Process. Journal of Innovative Image Processing. 2026;8(3):1076-1095. doi:10.36548/jiip.2026.3.016
BibTeX
@article{Midhunchakkaravarthy2026,
author = {Midhunchakkaravarthy, Divya and G., Muni Nagamani and V., Lakshman Narayana},
title = {Dual-Model Transfer Learning with Graph Neural Networks and SHAP Explainability for Robust Colorectal Tumor Margin Detection Process},
journal = {Journal of Innovative Image Processing},
year = {2026},
volume = {8},
number = {3},
pages = {1076--1095},
doi = {10.36548/jiip.2026.3.016}
}
EUVIMED