The result prediction of fluorescence in situ hybridization for breast cancer patients based on machine learning and deep learning models: a multicenter study.

Abstract

OBJECTIVE: To develop and validate a multicenter ultrasound-based predictive model for fluorescence in situ hybridization (FISH) results in HER2 (2+) breast cancer patients, aiming to provide a convenient and cost-effective tool to support clinical decision-making. MATERIALS AND METHODS: In this retrospective multicenter study, 5,888 breast cancer patients from six institutions were included. Radiomics features were extracted from ultrasound images using PyRadiomics, and deep learning features were obtained using a Vision Transformer (ViT). Feature selection was performed using least absolute shrinkage and selection operator (LASSO) regression. Multiple machine learning models were developed, and their performance was evaluated with the area under the curve (AUC). The DeLong test was used for model comparison. RESULTS: The proportion of FISH-positive cases ranged from 9.7% to 20.0% across the six centers. The fusion model combining ViT and radiomics signatures consistently outperformed the individual models in the training, test, and all external validation cohorts. The AUCs of the fusion model were 0.887 in the training cohort, 0.799 in the test cohort, and 0.763, 0.796, 0.734, and 0.632 in the four external validation cohorts, respectively (all P < 0.05). CONCLUSION: The proposed ultrasound-based fusion model enables accurate prediction of FISH assay results in HER2 (2+) breast cancer patients and may serve as a reliable decision-support tool to reduce unnecessary FISH testing in clinical practice.