Radiomics versus frozen deep embeddings for early pCR prediction on longitudinal breast MRI

In 183 patients from public ACRIN 6698/BMMR2, early longitudinal breast MRI prediction of pathologic complete response was modest and comparable: handcrafted radiomics best AUROC 0.670; frozen deep embeddings best AUROC 0.672 and best AUPRC 0.472. Models remain exploratory.

• The validated cohort included 183 patients with matched T0 and T1 DCE breast MRI image-mask pairs and binary pathologic complete response labels.
• Handcrafted radiomics: best T1-only random forest with 10 selected features achieved AUROC 0.670 and AUPRC 0.405 on the independent test subset.
• Frozen deep embeddings: best by AUROC was DELTA-only random forest (AUROC 0.672, AUPRC 0.408); best by AUPRC was AVG-only logistic regression (AUROC 0.666, AUPRC 0.472). End-to-end paired deep models overfit and were not retained; authors state models require independent external validation.

Automated summary

RadPigeon summaries are original and for information only. They are not clinical advice.