Abstract PS3-06-13: Early Prediction of Pathologic Complete Response in Breast Cancer via Graph Network Modeling of Intratumoral and Peritumoral Dynamics on DCE-MRI

Abstract Background: Pathological complete response (pCR) is an important marker associated with outcomes in breast cancer patients receiving neoadjuvant chemotherapy (NAC). However, pCR is confirmed only after surgery through pathological examination, making it difficult to assess treatment efficacy early and adjust therapy in time. Therefore, it is necessary to modelling intratumoral and peritumoral dynamics to predict pCR in early stage. Methods: We collected pre-treatment and early-treatment (one-cycle) DCE-MRI scans from Center A (n=259) and ACRIN 6698 trial (n=227). Center A was divided into training and internal testing sets in 8:2 ratio, and ACRIN 6698 trial served as an external test sets. Tumor regions were segmented and expanded by 5 mm to define the peritumoral regions. To capture intratumoral information, each tumor was partitioned into 100 superpixels, and then were categorize into three habitats using k-means clustering. Then we constructed graph with ten nodes—tumor, peritumoral region, and three habitats at both pre-treatment and one-cycle timepoints. A three-layer Graph Convolutional Network (TriGCN) was employed to model interactions between nodes. Finally, the output was concatenated with clinicopathological characteristics to build an integrated model (TriGCN-C). Results: The results presented the area under the curve (AUC) of the receiver operating characteristic curve. The integrated model (TriGCN-C) demonstrated superior performance, achieving AUCs of 0.85 in the training set, 0.81 in the internal testing set, and 0.79 in the external testing set. These results outperformed those of the clinical model (AUC: 0.69-0.75), as well as the single-timepoint models (AUC: 0.66-0.76) in the testing sets. Although the single-timepoint models outperformed the integrated model in the training set, they suffered from overfitting in the testing sets. In contrast, the integrated model captured the dynamic changes of the tumor during treatment, enabling better generalization and demonstrating robust prediction. Conclusion: Our study shows that DCE-MRI from pre-NAC and early-NAC can predict pCR, assisting in early adjustments of therapeutic strategies. Graph-based models have proven effective in capturing the interactions among different tumor nodes. These findings emphasize the model’s promise for predicting treatment response in clinical practice. Citation Format: Y. Tan, S. Du, L. Zhang, Z. Liu, J. Tian. Early Prediction of Pathologic Complete Response in Breast Cancer via Graph Network Modeling of Intratumoral and Peritumoral Dynamics on DCE-MRI abstract. In: Proceedings of the San Antonio Breast Cancer Symposium 2025; 2025 Dec 9-12; San Antonio, TX. Philadelphia (PA): AACR; Clin Cancer Res 2026;32(4 Suppl):Abstract nr PS3-06-13.