Spatial Attention-Enhanced Encoder-Decoder Network for Accurate Segmentation of the Prostate's Transition Zone.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Accurate segmentation of prostate and its substructures is crucial for prostate cancer localization and characterization.
• Existing deep learning models face challenges in precise segmentation of the prostate transitional zone.

Purpose of the Study:

• To propose a Spatial Attention Residual U-Net (Spatial ResU-Net) deep learning network for enhanced segmentation of the prostate transitional zone.
• To leverage spatial attention modules and residual connections for improved feature extraction and information flow.

Main Methods:

• Development of a novel Spatial ResU-Net architecture incorporating spatial attention modules and residual connections.
• Training and evaluation of the proposed model against 8 state-of-the-art deep learning segmentation models.
• Quantitative comparison using metrics such as Sensitivity, Dice Score, Hausdorff distance, and Average surface distance.

Main Results:

• The Spatial ResU-Net demonstrated superior performance compared to 8 other deep learning models.
• Improvements observed included at least 1% increase in Sensitivity and Dice Score.
• Significant enhancements in accuracy were noted with at least 0.05 mm improvement in Hausdorff distance and 0.09 mm in Average surface distance.

Conclusions:

• The proposed Spatial ResU-Net is a highly effective deep learning model for prostate transitional zone segmentation.
• This method offers improved accuracy and reliability for prostate cancer diagnosis and treatment planning.
• Spatial attention and residual connections are key components for advancing medical image segmentation tasks.