A machine learning model that classifies breast cancer pathologic complete response on MRI post-neoadjuvant...

Area of Science:

• Radiology and Medical Imaging
• Oncology
• Machine Learning in Medicine

Background:

• Pathologic complete response (pCR) in breast cancer after neoadjuvant chemotherapy (NAC) is crucial but cannot be assessed non-invasively.
• Current practice necessitates surgery for all patients to determine pCR, regardless of treatment response.

Purpose of the Study:

• To develop and validate a radiomics classifier for predicting breast cancer pCR on MRI after NAC.
• To enable non-invasive assessment of treatment response prior to surgical intervention.

Main Methods:

• Retrospective analysis of breast cancer patients (2014-2016) with pre- and post-NAC MRI and pathology reports.
• Automated radiomics analysis including segmentation, feature extraction, and recursive feature elimination random forest (RFE-RF) machine learning.
• Two models were trained: radiomics only (Model 1) and radiomics with molecular subtype (Model 2), addressing class imbalance with synthetic minority oversampling technique.

Main Results:

• The study included 278 invasive breast cancers; Model 1 achieved an Area Under the Receiver Operating Characteristic Curve (AUROC) of 0.83 in the test set.
• Model 2 (radiomics + molecular subtype) achieved an AUROC of 0.78 in the test set, showing similar accuracy to Model 1.
• No significant differences in pCR or molecular subtype were observed between training and test sets.

Conclusions:

• A validated radiomics classifier integrating radiomics features and molecular subtypes can accurately predict pCR on MRI post-NAC.
• This approach holds potential for non-invasive assessment of treatment response in breast cancer patients undergoing NAC.