MicroSegNet: A deep learning approach for prostate segmentation on micro-ultrasound images.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Biomedical Engineering

Background:

• Micro-ultrasound (micro-US) offers high-resolution imaging for prostate cancer diagnosis.
• Accurate prostate segmentation is vital for clinical applications but challenging on micro-US due to image artifacts and unclear boundaries.
• Existing segmentation methods struggle with the complexities of micro-US prostate imaging.

Purpose of the Study:

• To develop and evaluate MicroSegNet, a novel deep learning model for accurate prostate segmentation on micro-ultrasound images.
• To address the challenges of artifacts and indistinct borders in micro-US prostate segmentation.
• To improve the reliability of prostate segmentation for diagnosis, biopsy, and treatment planning.

Main Methods:

• Development of MicroSegNet, a multi-scale annotation-guided transformer UNet model.
• Introduction of an annotation-guided binary cross entropy (AG-BCE) loss function to focus on difficult-to-segment regions.
• Integration of AG-BCE loss with multi-scale deep supervision for enhanced feature learning.

Main Results:

• MicroSegNet achieved a Dice coefficient of 0.939 and a Hausdorff distance of 2.02 mm on a dataset of 75 patients (55 training, 20 testing).
• The model significantly outperformed several state-of-the-art segmentation methods.
• MicroSegNet demonstrated superior performance compared to three human annotators with varying experience levels.

Conclusions:

• MicroSegNet provides a robust and accurate solution for prostate segmentation in micro-ultrasound imaging.
• The proposed AG-BCE loss and multi-scale deep supervision effectively improve segmentation performance in challenging cases.
• This advancement holds significant potential for low-cost, accurate prostate cancer diagnosis and management.