Shadow-Consistent Semi-Supervised Learning for Prostate Ultrasound Segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Prostate segmentation in transrectal ultrasound (TRUS) images is critical for clinical procedures.
• Low image quality and shadow artifacts present significant challenges in accurate prostate segmentation.
• Existing methods struggle with the inherent difficulties of TRUS imaging.

Purpose of the Study:

• To develop a robust method for prostate segmentation in TRUS images.
• To overcome limitations posed by image quality and shadow artifacts.
• To enhance the efficiency of prostate segmentation through semi-supervised learning.

Main Methods:

• Proposed a Shadow-consistent Semi-supervised Learning (SCO-SSL) method.
• Introduced Shadow Augmentation (Shadow-AUG) to simulate and incorporate shadow artifacts.
• Implemented Shadow Dropout (Shadow-DROP) to enforce boundary inference from shadow-free regions.

Main Results:

• The proposed method significantly outperformed state-of-the-art approaches in the fully-supervised setting.
• SCO-SSL achieved highly competitive performance even with only 20% labeled data in the semi-supervised setting.
• Experiments were validated on two large clinical datasets (1,761 and 662 TRUS volumes).

Conclusions:

• The SCO-SSL method demonstrates superior performance and robustness in prostate segmentation.
• The approach shows significant clinical value by reducing the need for extensive data annotation.
• This work offers a promising solution for improving automated prostate segmentation in clinical practice.