A new AI framework can predict whether a breast cancer patient will respond completely to neoadjuvant chemotherapy — before the treatment course is finished.
Researchers published a study describing a 3D spatio-temporal model that feeds sequences of dynamic contrast-enhanced MRI scans into a graph neural network. The model tracks how a tumor changes across multiple imaging sessions, then estimates the likelihood of pathological complete response (pCR) — the absence of residual invasive cancer after chemotherapy ends. Tested on 585 patients from the public ISPY-2 dataset, the model outperformed both standard vision and self-supervised learning baselines across several classification metrics. The team also plans to release the code and a PyPI library for dataset curation when the paper is formally published.
Predicting pCR matters because it is one of the most reliable proxies for long-term survival in breast cancer — patients who achieve it tend to fare significantly better. Most current clinical decisions rely on static snapshots or clinical intuition; a model that reads the trajectory of tumor change over time gives oncologists a more dynamic signal to act on. That could mean earlier pivots to alternative therapies for patients unlikely to respond.
The method is still research-stage and was tested on a single public dataset, so real-world generalizability remains an open question — a familiar caveat in medical AI that has burned clinicians before.