Longitudinal MRI-Driven Multi-Modality Approach for Predicting Pathological Complete Response and B Cell Infiltration in Breast Cancer

Abstract

Accurately predicting pathological complete response (pCR) to neoadjuvant treatment (NAT) in breast cancer remains challenging due to tumor heterogeneity. This study enrolled 2279 patients across 12 centers and develops a novel multi-modality model integrating longitudinal magnetic resonance imaging (MRI) spatial habitat radiomics, transcriptomics, and single-cell RNA sequencing for predicting pCR. By analyzing tumor subregions on multi-timepoint MRI, the model captures dynamic intra-tumoral heterogeneity during NAT. It shows superior performance over traditional radiomics, with areas under the curve of 0.863, 0.813, and 0.888 in the external validation, immunotherapy, and multi-omics cohorts, respectively. Subgroup analysis shows its robustness across varying molecular subtypes and clinical stages. Transcriptomic and single-cell RNA sequencing analysis reveals that high model scores correlate with increased immune activity, notably elevated B cell infiltration, indicating the biological basis of the imaging model. The integration of imaging and molecular data demonstrates promise in spatial habitat radiomics to monitor dynamic changes in tumor heterogeneity during NAT. In clinical practice, this study provides a noninvasive tool to accurately predict pCR, with the potential to guide treatment planning and improve breast-conserving surgery rates. Despite promising results, the model requires prospective validation to confirm its utility across diverse patient populations and clinical settings.

Keywords: artificial intelligence; breast cancer; medical imaging; multi‐omics analysis; neoadjuvant treatment.

Figure 1

The flowchart of patient inclusion and exclusion in this study. Four cohorts: training (n = 431), external validation (n = 1595), immunotherapy (n = 88), and multiomics cohort (n = 165) are used for artificial intelligence model construction, independent validation, and biological interpretability analysis. NAT = neoadjuvant treatment, pCR = pathological complete response, MRI = magnetic resonance imaging.

Figure 2

Overview of the study design and workflow. A) Breast cancer patients from 12 institutions were divided into a training cohort (n = 431) and three validation cohorts (n = 1825). Pre‐NAT and mid‐NAT MRI scans were obtained, followed by histological analysis after surgery. B) Tumor subregions were analyzed for spatial heterogeneity before and after NAT. C) Feature extraction included intensity, shape, texture, and wavelet features, followed by ICC reproducibility tests, Mann‐Whitney U tests, Spearman analysis, and LASSO for feature selection. Models were constructed and compared. D) Clinical validation, transcriptomic, and scRNA‐seq analysis were performed to evaluate the model's biological relevance. NAT = neoadjuvant treatment, pCR = pathological complete response, MRI = magnetic resonance imaging, ICC = intraclass correlation coefficient, LASSO = Least Absolute Shrinkage and Selection Operator.

Figure 3

Model performance of artificial intelligence models in different cohorts. A–D) ROC curves for the training, external validation, neoadjuvant immunotherapy, and multi‐omics cohorts, respectively, showing the AUC for various models. E–H) Calibration curves for the training and validation cohorts comparing predicted probabilities and actual outcomes for Clin‐SHR and Clin‐TR models, demonstrating the models' predictive accuracy. NAT = neoadjuvant treatment, SHR = spatial habitat radiomics, TR = traditional radiomics, AUC = the area under the curve.

Figure 4

Spearman correlation coefficient matrices of selected imaging features and confusion matrices of predictive models. A,B) Spearman correlation coefficient matrices of selected radiomic features in the training and validation cohorts. C–F) Confusion matrices for the SHR model in the training, external validation, immunotherapy, and multi‐omics cohorts. G–J) Confusion matrices for the TR model in the training, external validation, immunotherapy, and multi‐omics cohorts. These results demonstrate the consistency of models' predictive performance and the radiomic features across different cohorts. pCR = pathological complete response.

Figure 5

The associations between radiomics feature and immune landscape of low and high Clin‐SHR groups using transcriptomic and scRNA‐seq analysis. A) Gene set enrichment analysis highlights significant differences in biological processes, such as chromosome organization and immune response, between the two groups. B) A heatmap shows correlation between radiomics features and various immune cell populations. Each cell type is evaluated for its normalized correlation with specific radiomics features, indicating potential links between radiomics patterns and the tumor immune microenvironment. C) A UMAP visualization displays the distribution of different cell types, color‐coded for epithelial, basal, endothelial, fibroblast, myeloid, T cell, and B cell populations. D) The expression of key marker genes in each cell type, showing their average expression levels in the overall population. E) A UMAP visualization displays the distribution of different cell types, color‐coded for epithelial, basal, endothelial, fibroblast, myeloid, T cell, and B cell populations in low and high Clin‐SHR groups.

Figure 6

The difference of immune cell composition in high versus low Clin‐SHR scores using transcriptomic and scRNA‐seq analysis. A,B) The proportions of cell types between low and high Clin‐SHR patients. C,D) The ratio of observed over expected cell numbers (Ro/e) of different cell types between low and high Clin‐SHR patients. E) The proportion of B cell infiltration, with significant differences observed between low and high Clin‐SHR groups (P = 0.029). F) Differential gene expression profiles across cell types, with B cell‐specific genes prominently upregulated in high Clin‐SHR patients.

Figure 7

A comprehensive analysis of B cell infiltration in low and high Clin‐SHR breast cancer. A) Six representative patients’ pre‐NAT and mid‐NAT MRI images, alongside UMAP visualization depicting the cellular composition of their tumors. The pie charts detail the proportional representation of various cell types. B) The GO‐BP terms were enriched in highly variable genes (HVGs) of B cells, with the most significant pathways related to B cell receptor signaling and various immune response processes. C) The significant difference in B cell scores between low and high Clin‐SHR groups, with high Clin‐SHR patients showed higher B cell scores (p = 0.00039).