Weakly supervised learning-based 3D bladder reconstruction from 2D ultrasound images for bladder volume measurement

Area of Science:

• Medical Imaging
• Radiation Oncology
• Artificial Intelligence

Background:

• Accurate bladder volume measurement is crucial for consistent patient anatomy in pelvic tumor radiation therapy.
• Traditional 2D ultrasound methods yield inaccurate bladder volume measurements due to anatomical variations.

Purpose of the Study:

• To enhance the precision of bladder volume measurement from 2D ultrasound images for patients undergoing pelvic tumor radiation therapy.

Main Methods:

• A weakly supervised learning-based 3D bladder reconstruction neural network was developed using retrospective ultrasound data from 130 patients.
• A novel loss function incorporating supervised segmentation and bladder compactness was designed.
• Bladder volume was calculated via voxel counting, with accuracy assessed using Dice Similarity Coefficient (DSC) and relative standard deviation (RSD) against CT scans.

Main Results:

• The method achieved a mean DSC of 0.94 and a mean absolute RSD of 6.3% with 12 ultrasound images.
• Even with only two images, a mean DSC of 0.90 and mean absolute RSD of 9.0% were obtained.
• This represents a significant improvement over commercial bladder scanners (mean absolute RSD of 13.6%).

Conclusions:

• The proposed 3D reconstruction method substantially improves bladder volume measurement accuracy.
• This technique shows significant potential for integration into future bladder volume measurement devices.