Predicting response to neoadjuvant therapy in breast cancer using longitudinal DCE-MRI deep learning integrated with
Lan Yan1, Xianming Huang2, Lan Liu1
1Department of Radiology, Jiangxi Cancer Hospital & Institute, Jiangxi Clinical Research Center for Cancer, The Second Affiliated Hospital of Nanchang Medical College, Nanchang, China.
Objective:
This study developed and validated a multimodal fusion model to enable the early and accurate prediction of pathological complete response (pCR) to neoadjuvant therapy (NAT) in breast cancer. The model integrates deep learning (DL) features derived from longitudinal dynamic contrast-enhanced magnetic resonance imaging (DCE-MRI) acquired early during treatment, peripheral blood inflammatory (PBI) indices, and baseline levels of tumor-infiltrating lymphocytes (TILs).
Methods:
A total of 262 breast cancer patients receiving NAT were retrospectively enrolled and divided into a training cohort (n=183) and a validation cohort (n=79) based on the time of surgery. Deep learning models (Pre-NAT DL and Post-2nd-NAT DL) were constructed using features extracted from pre-treatment (baseline) and post-second-cycle DCE-MRI images, respectively. An immune-inflammation model was built using baseline TILs and dynamically changing PBI indices. A clinical model was developed based on baseline clinicopathological characteristics. Finally, a combined model was constructed by integrating features from all the aforementioned modalities. The models were developed using various machine learning algorithms, and their predictive performance was assessed and compared.
Results:
In the validation cohort, the combined model achieved superior predictive performance, with an area under the receiver operating characteristic curve of 0.90 and specificity of 95%. Its performance was significantly better than that of any single-modality model. The Post-2nd-NAT DL model (AUC = 0.85) outperformed the Pre-NAT DL model (AUC = 0.75), confirming the critical predictive value of deep learning features from early-treatment DCE-MRI. The immune-inflammation model also exhibited independent predictive capability (AUC = 0.73).
Conclusion:
The combined model integrating deep learning features from early longitudinal DCE-MRI, dynamic systemic inflammatory indicators, and baseline TILs significantly enhances the early prediction of pCR to NAT in breast cancer. This multimodal fusion strategy offers a potential tool to aid personalized treatment planning in breast cancer patients undergoing NAT.