SAM-UNETR: Clinically Significant Prostate Cancer Segmentation Using Transfer Learning From Large Model

Prostate cancer (PCa) is one of the leading causes of cancer-related mortality among men worldwide. Accurate and efficient segmentation of clinically significant prostate cancer (csPCa) regions from magnetic resonance imaging (MRI) plays a crucial role in diagnosis, treatment planning, and monitoring of the disease, however, this is a challenging task even for the specialized clinicians. This study presents SAM-UNETR, a novel model for segmenting csPCa regions from MRI images. SAM-UNETR combines a transformer-encoder from the Segment Anything Model (SAM), a versatile segmentation model trained on 11 million images, with a residual-convolution decoder inspired by UNETR. The model uses multiple image modalities and applies prostate zone segmentation, normalization, and data augmentation as preprocessing steps. The performance of SAM-UNETR is compared with three other models using the same strategy and preprocessing. The results show that SAM-UNETR achieves superior reliability and accuracy in csPCa segmentation, especially when using transfer learning for the image encoder. This demonstrates the adaptability of large-scale models for different tasks. SAM-UNETR attains a Dice Score of 0.467 and an AUROC of 0.77 for csPCa prediction.

Automated summary

This study presents a novel model, SAM-UNETR, for segmenting clinically significant prostate cancer regions from MRI images. SAM-UNETR combines the features of Segment Anything Model and UNETR, and uses multiple image modalities and preprocessing steps to improve accuracy. Compared with three other models, SAM-UNETR achieves higher reliability and accuracy in csPCa segmentation.