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

Background/Objectives: Early prediction of pathologic complete response (pCR) during neoadjuvant therapy could improve diagnostic assessment of treatment response, adaptive treatment monitoring, and early prognostic stratification in breast cancer, but longitudinal public MRI archives are not immediately ready for reproducible paired-image modeling. The aim of this study was to develop a reproducible longitudinal breast MRI workflow from the public ACRIN 6698/BMMR2 archive and to compare handcrafted radiomics with frozen deep image embeddings for early treatment-response modeling and imaging-based pCR prediction. Methods: The raw archive was reduced to a validated 183-patient cohort with matched T0 and T1 cropped DCE image-mask pairs, preserved predefined train/test assignment, and binary pCR labels. Handcrafted radiomic features were extracted at T0 and T1 using a MIRP-based workflow, longitudinal delta descriptors were constructed, and train-only cleaning, filtering, and feature selection were applied before predictive modeling. In parallel, lesion-centered paired 2.5D tensors were generated for image-based deep analysis. Exploratory end-to-end paired deep-learning models were evaluated but showed substantial overfitting. Therefore, a frozen pretrained ResNet18 feature-extraction strategy was used to derive T0, T1, DELTA, and AVG image embeddings, which were then modeled using train-only selection and classical classifiers. Results: In the handcrafted radiomics branch, the strongest refined model was the T1-only random forest model with 10 selected features, achieving an AUROC of 0.670 and an AUPRC of 0.405 on the independent test subset. In the frozen deep-embedding branch, the best configuration by AUROC was the DELTA-only random forest model with 30 selected features (AUROC 0.672, AUPRC 0.408), whereas the strongest configuration by AUPRC was the AVG-only logistic regression model with 100 selected features (AUROC 0.666, AUPRC 0.472). End-to-end paired deep-learning models were technically feasible but were not retained because of marked overfitting. Conclusions: A complex public longitudinal breast MRI archive can be converted into a reproducible modeling cohort for early response assessment. Within this framework, handcrafted radiomics and frozen deep image embeddings yielded comparable exploratory discrimination but emphasized different representation spaces: the strongest handcrafted signal was concentrated in the early-treatment T1 phenotype, whereas the strongest deep signal was concentrated in latent longitudinal change and integrated paired-image state. These findings indicate that longitudinal breast MRI-derived biomarkers warrant further investigation for early treatment-response assessment; however, the present models remain exploratory and require independent external validation before any clinical application.