Integrating multimodal ultrasound imaging and machine learning for predicting luminal and non-luminal breast cancer subtypes.

Rationale and objectives: Breast cancer molecular subtypes significantly influence treatment outcomes and prognoses, necessitating precise differentiation to tailor individualized therapies. This study leverages multimodal ultrasound imaging combined with machine learning to preoperatively classify luminal and non-luminal subtypes, aiming to enhance diagnostic accuracy and clinical decision-making.

Methods: This retrospective study included 247 patients with breast cancer, with 192 meeting the inclusion criteria. Patients were randomly divided into a training set (134 cases) and a validation set (58 cases) in a 7:3 ratio. Image segmentation was conducted using 3D Slicer software, adhering to IBSI-standardized radiomics feature extraction. We constructed four model configurations-monomodal, dual-modal, trimodal, and four-modal-through optimized feature selection. These included monomodal datasets comprising 2D ultrasound (US) images, dual-modal datasets integrating 2D US with color Doppler flow imaging (CDFI) (US+CDFI), trimodal datasets incorporating strain elastography (SE) alongside 2D US and CDFI (US+CDFI+SE), and four-modal datasets combining all modalities, including ABVS coronal imaging (US+CDFI+SE+ABVS). Machine learning classifiers such as logistic regression (LR), support vector machines (SVM), AdaBoost (adaptive boosting), random forests(RF), linear discriminant analysis(LDA), and ridge regression were utilized.

Results: The four-modal model achieved the highest performance (AUC: 0.947, 95% CI: 0.884-0.986), significantly outperforming the monomodal model (AUC 0.758, ΔAUC +0.189). Multimodal integration progressively enhanced performance: trimodal models surpassed dual-modal and monomodal approaches (AUC 0.865 vs 0.741 and 0.758), and the four-modal framework showed marked improvements in sensitivity (88.4% vs 71.1% for monomodal), specificity (92.7% vs 70.1%), and F1 scores (0.905).

Conclusion: This study establishes a multimodal machine learning model integrating advanced ultrasound imaging techniques to preoperatively distinguish luminal from non-luminal breast cancers. The model demonstrates significant potential to improve diagnostic accuracy and generalization, representing a notable advancement in non-invasive breast cancer diagnostics.