New AI Improves Prostate Cancer Detection on MRI Scans
Better AI could mean fewer unnecessary biopsies and earlier cancer detection.
Researchers trained a prostate lesion segmentation model on biparametric MRI using a combined cohort of 1,579 lesion-containing training cases from five datasets—described as, to their knowledge, the largest collection of prostate lesion annotations used for this task. They investigated two improvements to nnU-Net: random sym4 (Symlet) wavelet augmentation outperformed random convolution augmentation in average Dice score, with only a modest F1 score reduction due to lower precision, and Connected-Component-based Dice (CC-Dice)-CE loss complemented BackSplit, performing better than BackSplit alone.
Together, these improvements reached a Dice score of 0.614 on the public Prostate158 test set, establishing a new state-of-the-art on that benchmark. Training code and model weights are publicly released.
- AI trained on 1,579 prostate MRI images from five hospitals sets a new accuracy record.
- Two simple tweaks to a standard AI framework made it better at outlining tumors.
- The AI and its code are free for other researchers to use and build upon.
Why It Matters
Better AI could lead to earlier prostate cancer detection and fewer unnecessary biopsies, saving lives and reducing anxiety.