Improving Prostate Gland Segmentation Using Transformer based Architectures
Transformer model achieves 0.902 Dice score, beating CNNs by up to 5% on noisy clinical data.
A research team has demonstrated that transformer-based AI models can significantly improve the accuracy and robustness of segmenting the prostate gland in MRI scans, a critical task for diagnosing and treating prostate cancer. The study, led by Shatha Abudalou, Yasin Yilmaz, and Yoganand Balagurunathan, rigorously compared the SwinUNETR and UNETR architectures against a conventional 3D UNet convolutional neural network (CNN). They trained and tested the models on a challenging dataset of 546 T2-weighted MRI volumes, annotated by two independent expert readers to account for real-world variability and label noise.
In their most successful configuration, the SwinUNETR model achieved an impressive average Dice Similarity Coefficient (DSC) of 0.902 when trained on a subset organized by gland size. This represents an improvement of up to five percentage points over the CNN baseline. The key finding is that the transformer's "global and shifted-window self-attention" mechanisms make it less sensitive to inconsistencies between different radiologists' annotations and variations across medical imaging sites. This ability to handle heterogeneity is a major hurdle for clinical AI adoption.
The research employed multiple training strategies—single cohort, 5-fold cross-validated mixed cohort, and gland size-based datasets—with hyperparameters optimized by the Optuna framework. SwinUNETR consistently outperformed UNETR and the traditional CNN, especially in the more complex, mixed-training scenarios designed to mimic clinical reality. The study concludes that SwinUNETR's combination of high accuracy and maintained computational efficiency makes it a strong candidate for robust, real-world clinical deployment, moving AI-assisted diagnostics from the lab to the clinic.
- SwinUNETR transformer model achieved a top Dice score of 0.902, beating a 3D UNet CNN by up to 5 percentage points on prostate MRI segmentation.
- Tested on 546 MRI volumes with annotations from two independent readers to simulate real-world label noise and inter-reader variability.
- Proves transformer architectures (specifically SwinUNETR) are more robust to clinical data heterogeneity than CNNs, a critical step for reliable medical AI.
Why It Matters
More reliable AI segmentation can improve prostate cancer diagnosis consistency, reduce radiologist workload, and accelerate treatment planning.