Pillar-0 + TDN predicts breast cancer pCR with 73.6% AUROC
A frozen 3D encoder plus temporal network beats static imaging for breast cancer response prediction.
Predicting whether breast cancer patients will achieve pathologic complete response (pCR) during neoadjuvant chemotherapy (NAC) is a critical clinical challenge. Most imaging models rely on a single static timepoint, missing the dynamic changes tumors undergo during treatment. A new paper from researchers including Fidel Omar Tito Cruz and colleagues tackles this with a longitudinal deep learning framework that tracks tumors across the entire treatment journey—from pre-treatment to pre-surgery.
The architecture combines Pillar-0, a frozen 3D foundation encoder, with a newly designed Temporal Dynamics Network (TDN). TDN fuses time-aware volumetric embeddings from four serial DCE-MRI scans with clinical and treatment metadata to predict pCR. Tested on a combined cohort of 982 patients from the I-SPY2 and ACRIN-6698 trials, the model achieved a test AUROC of 73.6% and balanced accuracy of 69.1%. Notably, while clinical variables alone provided the strongest signal, merging longitudinal 3D imaging with clinical data improved predictions, confirming that temporal imaging captures complementary information. The work was accepted at the AMAI Workshop at MICCAI 2026, and the team has released their code, offering a reproducible foundation for future research in treatment-response assessment.
- Longitudinal framework uses frozen Pillar-0 3D foundation encoder plus Temporal Dynamics Network over four DCE-MRI timepoints.
- Achieves 73.6% test AUROC and 69.1% balanced accuracy on 982 patients from I-SPY2 and ACRIN-6698.
- Fusing longitudinal imaging with clinical data improves pCR prediction; source code is public on arXiv.
Why It Matters
This enables earlier, more accurate breast cancer treatment response prediction, potentially guiding personalized neoadjuvant chemotherapy decisions.