Deep-learning based discrimination of pathologic complete response using MRI in HER2-positive and triple-negative...

Area of Science:

• Radiology and Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Accurate assessment of treatment response after neoadjuvant chemotherapy (NAC) is vital for breast cancer management.
• Current imaging techniques struggle to reliably differentiate between complete pathologic response and residual disease.
• Human epidermal growth factor receptor 2 (HER2)-positive and triple-negative breast cancers have distinct treatment pathways requiring precise response evaluation.

Purpose of the Study:

• To develop and validate deep learning models for predicting residual breast cancer after NAC.
• To evaluate the performance of models utilizing dynamic contrast-enhanced MRI (DCE-MRI) and clinical data.
• To compare the efficacy of different deep learning approaches, including phase-specific and whole-image analysis.

Main Methods:

• A 3D convolutional neural network (CNN) was trained on DCE-MRI and clinical data from 724 patients.
• The model was validated on an independent set of 128 patients with HER2-positive or triple-negative breast cancer.
• Performance was assessed by comparing models trained on early-phase, delayed-phase, and combined DCE-MRI data, as well as whole vs. cropped images.

Main Results:

• The delayed-phase deep learning model achieved a superior area under the receiver operating characteristic curve (AUC) of 0.74, outperforming the early-phase model (AUC=0.69) and a combined model (AUC=0.70).
• Models incorporating multiple dynamic phases and clinical data showed statistically significant improvements over single-phase models.
• Deep learning models using uncropped whole MRI images demonstrated significantly lower performance (AUCs 0.45-0.54).

Conclusions:

• Deep learning models, particularly those utilizing delayed-phase DCE-MRI, show potential for improving the accuracy of residual breast cancer detection post-NAC.
• The findings suggest that focused analysis of specific MRI phases is more effective than whole-image analysis for this task.
• Further research including external validation is recommended to enhance model generalizability and clinical utility.