Longitudinal Breast MRI for Early Treatment-Response Modeling: A Comparative Study of Handcrafted Radiomics and...

Area of Science:

• Medical Imaging
• Radiology
• Machine Learning in Oncology

Background:

• Early prediction of pathologic complete response (pCR) is crucial for optimizing breast cancer treatment and patient stratification.
• Longitudinal breast MRI data are essential for monitoring treatment response but require reproducible workflows.
• Public archives like ACRIN 6698/BMMR2 offer valuable data but need careful curation for modeling.

Purpose of the Study:

• To develop a reproducible longitudinal breast MRI workflow from the ACRIN 6698/BMMR2 archive.
• To compare handcrafted radiomics with frozen deep image embeddings for early treatment-response modeling.
• To predict imaging-based pCR in breast cancer patients undergoing neoadjuvant therapy.

Main Methods:

• A cohort of 183 patients with matched T0 and T1 DCE MRI scans and pCR labels was curated.
• Handcrafted radiomic features were extracted and analyzed using longitudinal delta descriptors.
• Frozen ResNet18 embeddings were generated for T0, T1, DELTA, and AVG states and modeled using classical classifiers.

Main Results:

• Handcrafted radiomics achieved an AUROC of 0.670 with a T1-only random forest model.
• Frozen deep embeddings yielded comparable results, with a DELTA-only model achieving an AUROC of 0.672.
• The strongest deep learning signal was observed in longitudinal changes and integrated image states, while radiomics highlighted the T1 phenotype.

Conclusions:

• A reproducible longitudinal breast MRI modeling cohort can be established from public archives.
• Handcrafted radiomics and deep image embeddings offer comparable, albeit distinct, insights into treatment response.
• Longitudinal MRI biomarkers show promise for early treatment response assessment, necessitating external validation for clinical use.