UniPosSeg : A Unified Framework for Dual-Position Breast Tumor Segmentation in MRI via Velocity-Field Registration Augmentation and Wavelet-Residual U-Net

Abstract

Purpose: In the clinical management of breast cancer, standard diagnostic imaging is typically performed in the prone position, whereas breast-conserving surgery and radiotherapy require the patient to be in the supine position. This positional transition induces a complex compound domain shift: the change in gravitational orientation leads to substantial non-rigid spatial deformation of the breast tissue, while the switching of radiofrequency (RF) coils results in significant variations in image intensity and contrast. This study aims to develop a unified, end-to-end deep learning framework (UniPosSeg) to simultaneously address spatial deformation and intensity variations without the requirement for pre-computed, perfectly registered paired data, thereby facilitating robust and automated segmentation of breast magnetic resonance imaging (MRI) tumors across dual positions.Materials and Methods: This study was based on a prospectively collected clinical cohort comprising 221 patients with a total of 442 dynamic contrast-enhanced (DCE) MRI scans (one prone and one supine scan per patient). The UniPosSeg framework approaches the aforementioned challenges through three core modules. First, a Velocity-field Registration Augmentation (VRA) module is introduced to model the substantial positional deformation as continuous fluid motion, generating physically plausible intermediate transition images to augment the spatial diversity of the training data. Second, a Feature-wise Linear Modulation (FiLM) module is integrated into the network to dynamically adjust the scaling and shifting of feature maps based on input positional information, compensating for the image intensity variations caused by RF coil switching. Finally, a Wavelet Residual Convolution (WRC) module is designed to extract high-frequency edge textures using orthogonal decomposition in the frequency domain, mitigating the loss of boundary information for small tumors in deep network layers.Results: Evaluated on a combined test set containing dual-position images, UniPosSeg demonstrated improved performance compared to classical U-Net variants, the medical segmentation benchmark nnU-Net, and recent Mamba-based architectures. Quantitative assessments showed that the proposed framework achieved a Dice Similarity Coefficient (DSC) of 0.6650 and reduced the 95\% Hausdorff Distance ($\text{HD}_{95}$) to 59.83 mm. Ablation studies confirmed the independent contributions and complementarity of the VRA, FiLM, and WRC modules.Conclusion: This study indicates that explicitly decoupling and independently processing spatial deformation and intensity variations within a unified deep learning architecture can effectively mitigate the technical challenges of cross-position tumor segmentation. By reducing the reliance on external image registration, the UniPosSeg framework offers a potential automated auxiliary tool for preoperative planning and radiotherapy target delineation in breast cancer, helping to bridge the geometric discrepancy between diagnostic imaging and physical intervention.