Automatic prostate segmentation using deep learning on clinically diverse 3D transrectal ultrasound images.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Oncology

Background:

• Manual prostate segmentation in 3D transrectal ultrasound (TRUS) is time-consuming and challenging during intraoperative procedures.
• Accurate segmentation is crucial for effective needle guidance in prostate cancer diagnosis and treatment.

Purpose of the Study:

• To develop a supervised deep learning algorithm for generalizable 3D prostate segmentation using diverse 3D TRUS images.
• To create a rapid and accurate segmentation method to improve workflow efficiency for needle-based prostate cancer procedures.

Main Methods:

• A modified 2D U-Net was developed, predicting segmentation on radially sampled slices and reconstructing a 3D surface.
• The model was trained and validated on 3D TRUS images from various acquisition methods and machines.
• Performance was evaluated against 3D networks using metrics like Dice Similarity Coefficient (DSC), Mean Surface Distance (MSD), and Hausdorff Distance (HD).

Main Results:

• The proposed method achieved high accuracy with a median DSC of 94.1%, median MSD of 0.89 mm, and median HD of 2.89 mm.
• It demonstrated significant improvement over optimized 3D networks across most evaluated metrics.
• Segmentation time was under 0.7 seconds per prostate, suitable for intraoperative use.

Conclusions:

• The developed deep learning algorithm provides a fast, accurate, and generalizable intraoperative solution for 3D prostate segmentation in TRUS images.
• This method can decrease procedure times and support the growing use of 3D TRUS in needle-based prostate cancer interventions.