Ultrasound prostate segmentation based on multidirectional deeply supervised V-Net.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Transrectal ultrasound (TRUS) is crucial for real-time prostate cancer interventions like biopsy and brachytherapy.
• Accurate prostate segmentation is vital for treatment planning and execution.
• Manual segmentation is time-consuming and prone to observer variability.

Purpose of the Study:

• To develop a deep learning-based method for automatic prostate segmentation using TRUS.
• To integrate deep supervision into a 3D patch-based V-Net architecture.
• To overcome optimization challenges in training deep networks with limited data.

Main Methods:

• A multidirectional deep learning approach using a 3D V-Net with integrated deep supervision was developed.
• A hybrid loss function combining binary cross-entropy and Dice loss was employed for training.
• The method segments prostates from TRUS images using patch extraction, adaptive labeling, patch fusion, and contour refinement.

Main Results:

• The method was tested on 44 patients' TRUS images, demonstrating high accuracy.
• Mean Dice similarity coefficient (DSC) was 0.92 ± 0.03.
• Quantitative metrics including Hausdorff distance (HD), mean surface distance (MSD), and residual mean surface distance (RMSD) were favorable.

Conclusions:

• A novel, deeply supervised deep learning method for automatic TRUS prostate segmentation was successfully developed.
• The approach demonstrated clinical feasibility and accuracy comparable to manual segmentation.
• This technique shows promise as a valuable tool for prostate cancer diagnosis and therapy.