New PaRC-mix training slashes prostate cancer segmentation errors by 75%
Trained on one MRI scanner, tested on five—accuracy gap almost eliminated
A team of researchers (including Josiah Simeth, Sudharsan Madhavan, and others) has introduced a novel training technique called Parallel-Route Coherent Mixup Regularization (PaRC-mix) designed to make deep learning segmentation models work reliably across different MRI scanners—a critical barrier for MRI-guided adaptive radiotherapy (MRgART) in prostate cancer. Unlike standard mixup that blends only input images, PaRC-mix linearly combines features at multiple network layers, creating richer augmentations that help models ignore scanner-specific artifacts. The method was implemented on two architectures: a multiple resolution residual network (MRRN) and UNet++.
On a challenging multi-domain test set—1,547 prostate cancer samples from 3T Siemens, 3T Philips, and 1.5T Elekta Unity MR-Linac scanners—PaRC-mix-trained models significantly outperformed baseline and input-mix variants. The composite DSC, HD95, and MSD score showed the accuracy gap between aggressive and non-aggressive lesions shrank from 21.1 to 5.2 for MRRN and from 19.5 to 7.9 for UNet++. This network-agnostic, easy-to-implement approach could accelerate adoption of MRgART by eliminating the need for site-specific retraining, directly translating to safer, more personalized radiation therapy.
- PaRC-mix applies feature mixup at multiple network layers, not just input, boosting generalization across scanner manufacturers.
- Models trained on 2,029 GE 3.0T samples tested on 1,547 samples from Siemens, Philips, and Elekta scanners.
- Accuracy gap between lesion types reduced by up to 75% (from 21.1 to 5.2 composite score) compared to no mixup training.
Why It Matters
Enables one-model-fits-all prostate cancer segmentation across MRI systems, streamlining adaptive radiotherapy workflows globally.