Deep Attentive Features for Prostate Segmentation in 3D Transrectal Ultrasound

Area of Science:

• Medical imaging
• Artificial intelligence
• Computer-aided diagnosis

Background:

• Accurate prostate segmentation in transrectal ultrasound (TRUS) is crucial for image-guided interventions and treatment planning.
• Challenges include ambiguous boundaries, inhomogeneous intensity, and variable prostate shapes in TRUS images.
• Existing automatic segmentation methods struggle with these inherent difficulties.

Purpose of the Study:

• To develop a novel 3D deep neural network with attention modules for enhanced prostate segmentation in TRUS images.
• To leverage complementary information across different convolutional neural network (CNN) layers for improved segmentation accuracy.
• To address the limitations of current methods in handling prostate shape variability and image quality issues.

Main Methods:

• A novel 3D deep neural network architecture incorporating attention modules was developed.
• The attention module selectively integrates multi-level features from different CNN layers.
• This refines features by suppressing noise in shallow layers and enhancing details in deep layers.

Main Results:

• The proposed method demonstrated satisfactory segmentation performance on challenging 3D TRUS volumes.
• Experimental results confirmed the effectiveness of the attention mechanism in improving segmentation accuracy.
• The approach successfully addressed issues related to ambiguous boundaries and intensity variations.

Conclusions:

• The developed 3D deep neural network with attention modules offers a robust solution for automatic prostate segmentation in TRUS.
• The attention mechanism provides a general strategy for aggregating multi-level deep features, applicable to other medical imaging tasks.
• This work contributes to advancing image-guided prostate interventions and treatment planning through improved automated segmentation.