MRI radiomics models show moderate accuracy for predicting HER2 status in young breast cancer patients

ObjectiveHabitat imaging can quantify intratumoral heterogeneity in young breast cancer patients, providing support for the prediction of HER2 expression levels. Therefore, this study aimed to compare the predictive ability of habitat models and conventional whole-tumor models for HER2 expression status in young breast cancer patients using multiparametric MRI.MethodsA retrospective cohort consisting of 375 young breast cancer patients (age < 40 years) who underwent preoperative MRI scanning at two medical centers was included in this study. Two binary classification tasks were designed: Task 1 (HER2 negative expression vs. HER2 positive expression) and Task 2 (HER2 zero expression vs. HER2 low expression). The training cohort (n=206) comprised patients from Center 1, with the external validation cohort (n=169) recruited from Center 2. Clinicopathological and MRI characteristics were collected. Radiomics features based on whole-tumor and habitat regions were extracted from DCE-MRI and DWI images, respectively. Clinical models, conventional whole-tumor models, habitat models, and combined models were constructed. Subsequently, model performance was evaluated by the AUC, sensitivity, and specificity.ResultsIn Task 1, the AUC of the clinical model, conventional whole-tumor model, habitat model, and combined model in the training cohort were 0.683, 0.731, 0.761, and 0.768 respectively. In Task 2, no clinicopathological features were determined as independent risk factors, thus no clinical model was developed. The AUC for the whole-tumor model, habitat model, and combined model in the training cohort were 0.673, 0.649, and 0.758 respectively.ConclusionThe habitat model exhibited better discriminatory effectiveness in identifying HER2 positive expression in young breast cancer patients, in comparison to the whole-tumor radiomics model. The integration of conventional whole-tumor radiomics features with habitat features and clinicopathological characteristics can enhance model performance.