Radiopathomic Graph Deep Learning for Multiscale Spatial-Contextual Modeling of Intratumoral Heterogeneity to Predict Breast Cancer Response to Neoadjuvant Therapy.
Authors
Affiliations (4)
Affiliations (4)
- Imaging Center, Harbin Medical University Cancer Hospital, Haping Road No. 150, Nangang District, Harbin 150081, China.
- Department of Pathology, Harbin Medical University Cancer Hospital, Harbin, China.
- Scientific Research Center, Fourth Affiliated Hospital of Harbin Medical University, Harbin, China.
- CREATIS, CNRS UMR 5220-INSERM U1206-University Lyon 1-INSA Lyon-University Jean Monnet Saint-Etienne, Lyon, France.
Abstract
Purpose To develop an explainable radio-pathomic graph deep-learning (RPGDL) system for multiscale spatial-contextual modeling of intratumoral heterogeneity (ITH) and evaluate its performance for the prediction of pathologic complete response (pCR) to neoadjuvant therapy (NAT) in breast cancer (BC). Materials and Methods The RPGDL system was developed from dual-center retrospective analysis of patients with biopsy-proven invasive BC (May 2018-August 2024). For each tumor, individual radiomic and pathomic graphs were generated from pretherapeutic MRI and hematoxylin and eosin-stained biopsy slide images, respectively. These graphs were then processed by three distinct graph neural networks (GNNs): radiomic, pathomic, and radio-pathomic GNNs. GNN performance was assessed with the area under the receiver operating characteristic curve (AUC), net reclassification index (NRI), and integrated discrimination improvement (IDI). A multifaceted approach was used to explain the GNNs' predictions. Results The training/external test set included 582/468 patients from Centers I/II (mean ages, 52 ± 9/50 ± 10 years). The radiomic GNN achieved AUCs of 0.89 (95% CI 0.85-0.93, training) and 0.84 (95% CI 0.80-0.89, external test); the pathomic GNN achieved AUCs of 0.87 (95% CI 0.83-0.91, training) and 0.83 (95% CI 0.78- 0.88, external test), with no significant difference between them (<i>P</i> >.05). The radio-pathomic GNN outperformed both single-modality GNNs (AUC [95% CI], 0.95 [0.92-0.98]/0.91 [0.87-0.94]; training/external test; <i>P</i> <.05; NRI and IDI confirmed). Pathomic graphs dominated probability increases for pCR predictions; while radiomic graphs drove probability decreases for non-pCR predictions. Multifaceted analyses verified GNNs' explainability. Conclusion The developed RPGDL system enabled multiscale spatial-contextual ITH modeling for highperformance, explainable prediction of pCR to NAT in BC. ©RSNA, 2026.