Machine Learning-Based Breast Deformation Estimation for MRI-Mammography Registration In Dense Breasts: Craniocaudal and Mediolateral Oblique Views
Abstract
Purpose
Dense breast tissue substantially reduces mammography sensitivity, posing diagnostic challenges particularly among Asian women (up to 50%). Multimodal registration between breast MRI and mammography improves lesion localization; however, finite element method (FEM)–based deformation modeling is computationally expensive and limits clinical scalability under large compression. This study develops a machine learning (ML)-based approach to approximate FEM-driven breast deformation, enabling efficient MRI-mammography registration under clinical large-deformation conditions for both craniocaudal (CC) and mediolateral oblique (MLO) views.
Methods
A validated FEM framework was used as ground truth to simulate mammographic compression exceeding 100 mm for CC and MLO views. FEM-derived node-wise displacements and physics-informed features were used to train an XGBoost model to predict three-dimensional deformation. Model performance was evaluated using a strict leave-one-patient-out validation scheme, with emphasis on final compression state corresponding to mammographic acquisition. In addition to global deformation accuracy, tumor-specific displacement errors were analyzed to assess lesion localization performance.
Results
At the final compression step (patient-level), mean displacement errors were 11.22 mm for CC and 7.87 mm for MLO views, with corresponding root mean square errors of 6.04 mm and 4.28 mm. The larger error observed in CC view reflects strong non-linearity introduced by extreme compression (>100 mm). In contrast, MLO demonstrated more stable deformation behavior and reduced error. Tumor regions exhibited slightly higher displacement errors than non-tumor tissue, attributable to mechanically heterogeneous interactions within internal breast structures; nevertheless, tumor deformation followed consistent physics-informed trends captured by the ML model.
Conclusion
This study demonstrates that ML can approximate FEM-based breast deformation with substantially reduced computation. By providing the first systematic evaluation of MLO-view deformation and explicitly analyzing tumor-specific displacement, this work addresses key gaps in prior studies. The proposed framework supports scalable, coarse-to-fine MRI–mammography registration in dense breasts, representing a practical step toward clinically deployable multimodal imaging workflows.