Habitat-Based Multimodal Deep Learning for Preoperative Prediction of Axillary Lymph Node Metastasis in Breast Cancer: A Two-Center Multiparametric MRI Study
Background : Axillary lymph node (ALN) status is a critical determinant of breast cancer staging and treatment planning. This study aimed to develop and validate a habitat-based multimodal deep learning (HMDL) model that learns whole-tumor and habitat subregion-level multiparametric MRI features, radiomics features, and clinical features within a unified network to enable noninvasive preoperative prediction of ALN metastasis in breast cancer. Methods : This retrospective study enrolled 201 patients with breast cancer from two centers. At Center 1, 160 patients were randomly divided into training (n = 128) and internal validation (n = 32) sets at the patient level in an 8:2 ratio. The remaining 41 patients from Center 2 served as an independent external validation set. A two-stage 3D U-Net was used for automatic segmentation of the breast tumor region of interest (ROI). Based on dynamic contrast-enhanced MRI (DCE), fat-suppressed T2-weighted imaging (T2FS), and diffusion-weighted imaging (DWI), K-means clustering was applied to partition intratumoral voxels into habitat subregions, and radiomics features were extracted from the whole tumor and each subregion. The HMDL model used a 2D convolutional neural network (CNN) as the backbone, extracting whole-tumor and habitat subregion image features along three orthogonal planes, fusing them via cross-modal and cross-region multi-head self-attention into a unified imaging representation, and concatenating it with radiomics and clinical features for classification. Ablation experiments were performed to evaluate the contribution of each information source, input sequence, and input strategy. Clinical and clinical-habitat radiomics models were constructed as comparators, and DeLong tests were used to compare the area under the receiver operating characteristic curve (AUC) among the three models. Results : The automatic segmentation model achieved Dice similarity coefficients of 0.842 and 0.804 on the internal and external validation sets, respectively. Based on the Calinski-Harabasz index, each tumor was partitioned into 3 habitat subregions. Ablation experiments showed that removing habitat subregion representations, radiomics features, clinical features, or any single sequence decreased the AUC on both validation sets. Among the three sequences, removal of DWI caused the largest AUC decrease in the HMDL (w/o DWI) model (AUC, 0.702 and 0.676, respectively), with statistically significant differences (P = 0.015, P = 0.012). The HMDL model achieved AUCs of 0.929 and 0.913 on the internal and external validation sets, respectively, both higher than the clinical-habitat radiomics model (0.869 and 0.841) and the clinical model (0.736 and 0.709). The difference between the HMDL and clinical models was statistically significant on the external validation set (P = 0.027). Conclusions : The HMDL model enables noninvasive quantitative prediction of ALN metastasis in breast cancer by jointly optimizing whole-tumor and habitat subregion-level imaging information, radiomics features, and clinical features within a unified network end-to-end, and demonstrates robust performance in independent external validation, offering a novel technical approach for preoperative ALN assessment.
Authors
- Liyou Huang
- Xu Gao (ORCID: https://orcid.org/0000-0002-5155-835X)
- Pei Cai
- Jiahai Huang
- Quanquan Guo
- Xun Wang
Institutions
- Xuzhou Medical College (CN)
- Nanjing Medical University (CN)
Publication Details
- Journal
- Research Square
- Published
- 2026-11-19
- DOI
- https://doi.org/10.21203/rs.3.rs-11496762/v1
- Primary Topic
- Radiomics and Machine Learning in Medical Imaging
- Type
- preprint