Multimodal combined model integrating 2.5D deep learning and habitat radiomics for malignancy discrimination in ≤2 cm BI-RADS 4 breast lesions.

Background: Differentiating between ≤2 cm Breast Imaging Reporting and Data System (BI-RADS) category 4 breast lesions remains challenging and often leads to unnecessary biopsies. This study aimed to develop and externally validate a multimodal fusion model to improve malignancy discrimination in ≤2 cm BI-RADS 4 lesions. Methods: A total of 686 women with ≤2 cm BI-RADS 4 breast lesions were included in this dual-center retrospective study (The Third Xiangya Hospital, Central South University: n=526, training cohort, n=368, internal validation cohort, n=158; Changsha Central Hospital: n=160, external test cohort, n=160). Separate models based on automated breast volume scanner (ABVS) and strain elastography (SE) were developed, including ultrasound (US) models, radiomics-habitat (Rad-Habitat) models, and two-and-a-half dimensional (2.5D) and two-dimensional (2D) deep learning (DL) models. The predicted probabilities from the single-modality models were then integrated to construct two fusion models: the Combined (ABVS) model and the Combined (ABVS + SE) model. Model discrimination, calibration, and clinical utility were evaluated using the areas under the curve (AUCs), calibration curves, Brier scores, and decision curve analysis (DCA), with additional subgroup analyses of BI-RADS 4a lesions. Results: In the internal validation and external test cohorts, the US models provided baseline discrimination (AUCs: 0.808-0.852). The Rad-Habitat models achieved AUCs ranging from 0.845-0.881, while the DL models achieved AUCs ranging from 0.874-0.887. The fusion models consistently outperformed all the single-modality models: the Combined (ABVS) model yielded AUCs of 0.948 and 0.942 in the internal and external cohorts, respectively, and the Combined (ABVS + SE) model further increased the internal AUC to 0.969 (DeLong test, P<0.05 for all pairwise comparisons vs. single-modality models). Both fusion models demonstrated good calibration with low Brier scores and achieved the highest net clinical benefit on DCA. In the ≤2 cm BI-RADS 4a subgroup, the combined models maintained high discriminatory performance (AUCs: 0.920-0.949) across the internal validation and external test cohorts. Conclusions: The multimodal fusion model integrating clinical US, habitat radiomics, and 2.5D DL significantly improved malignancy discrimination for ≤2 cm BI-RADS 4 lesions with robust external validation, and may reduce unnecessary biopsies while supporting individualized patient management.