Abstract
OBJECTIVE: This study aims to compare the potential performance of multimodal MRI-based models in discriminating TNBC. METHODS: Clinical and MRI data from 162 female breast cancer patients diagnosed at our hospital between March 2022 and June 2024 were retrospectively collected. Clinical-conventional MRI model, qMRI model, radiomics model, and three integrated models were constructed on the borderline SMOTE-processed training set. ROC and DCA analyses were used to evaluate the discrimination performance and clinical net benefit. The optimal model was selected by validating the stability of the integrated models on the test set. RESULTS: The mean age of 162 patients was 50.549 ± 9.563 years. No significant differences were found between the training set (n = 113) and the test set (n = 49) (all P > 0.05). The performance of models on the borderline SMOTE-processed training set was as follows: the clinical-conventional MRI model (including Ki-67 and BPE) achieved an AUC of 0.896; the qMRI model (including ADC(tunor) and T2-mapping (tumor)) achieved an AUC of 0.781; the radiomics model (including 9 key features) achieved an AUC of 0.905. The fully integrated model (clinical-conventional MRI model integrating all multimodal features) demonstrated optimal performance (AUC: 0.986). However, during independent validation in the test set, the fully integrated model's AUC decreased to 0.718, while the clinical-conventional MRI model maintained relatively stable discriminatory performance (AUC: 0.862). CONCLUSION: The clinical-conventional MRI model demonstrated stable discriminatory performance. While complex models show potential, their ability to generalize was limited by sample size, highlighting the need for validation with larger datasets.