New AI predicts cancer invasion from MRI without manual annotations
Teacher-student model learns from anatomy masks at train time, then works without them at inference.
Perineural invasion (PNI) is a key predictor of poor outcomes in intrahepatic cholangiocarcinoma, but it's currently confirmed only through post-surgical pathology. A team of researchers from multiple institutions has developed a new deep learning method that predicts PNI from preoperative T2‑weighted MRI scans, without requiring radiologist-defined variables, contrast agents, or manual annotations at inference time.
Their approach is a teacher-student distillation framework. During training, a teacher network sees both the MRI and anatomical masks (tumor and liver) to learn dense token routing—essentially learning which image patches are most informative. The student network, which never sees masks, is forced to retain and aggregate only a fixed budget of tokens by mimicking the teacher's routing decisions. The result: a lightweight model that runs in 8.02 ms per case on an NVIDIA Jetson Orin Nano Super (1.43 GFLOPs) and achieves a mean AUROC of 0.750 on a 155‑patient dataset. This makes it suitable for deployment in resource‑constrained clinical settings.
- Uses only T2-weighted MRI – no contrast agents or manual annotations required at inference
- Achieves 0.750 mean AUROC on 155 patients for perineural invasion prediction
- Runs in 8.02 ms per case on edge hardware (Jetson Orin Nano Super) with 1.43 GFLOPs
Why It Matters
Enables non‑invasive preoperative prediction of cancer invasion, reducing reliance on surgical pathology and manual radiologist inputs.