Structure boundary-preserving U-Net for prostate ultrasound image segmentation

Area of Science:

• Medical Imaging
• Computer Vision
• Biomedical Engineering

Background:

• Accurate prostate localization is critical for ultrasound-guided biopsies in prostate cancer diagnosis.
• Challenges include ambiguous prostate boundaries and imprecise manual delineations in ultrasound images.

Purpose of the Study:

• To develop a precise prostate segmentation method for improved localization in ultrasound images.
• To introduce a Boundary Structure-Preserving U-Net (BSP U-Net) that incorporates prostate shape priors.

Main Methods:

• Proposed a Boundary Structure-Preserving U-Net (BSP U-Net) integrating an active shape model-based key point selection module.
• Trained and evaluated the BSP U-Net on the PH2 + ISBI 2016 challenge dataset and a private prostate ultrasound dataset.

Main Results:

• Achieved a Dice Similarity Coefficient (DSC) of 95.94% and Jaccard Coefficient (JC) of 88.58% on the PH2 + ISBI 2016 dataset.
• Demonstrated high performance on prostate contouring with 97.05% pixel accuracy, 93.65% mean Intersection over Union, 92.54% DSC, and 93.16% JC.
• Outperformed existing state-of-the-art methods in prostate ultrasound image segmentation.

Conclusions:

• The BSP U-Net effectively preserves boundary structures and improves prostate segmentation accuracy.
• The proposed method offers a significant advancement for precise prostate localization in clinical ultrasound imaging.
• BSP U-Net shows superior performance compared to current state-of-the-art techniques.