Urinary bladder segmentation in CT urography using deep-learning convolutional neural network and level sets.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Urology

Background:

• Accurate bladder segmentation in CT urography (CTU) is crucial for computer-aided detection of bladder cancer.
• Traditional segmentation methods face challenges with strong boundaries and require significant user input.

Purpose of the Study:

• To develop and evaluate a computerized system for bladder segmentation in CTU using deep learning.
• To improve the accuracy and efficiency of bladder segmentation for bladder cancer detection.

Main Methods:

• A deep-learning convolutional neural network (DL-CNN) was trained on 160,000 regions of interest from CTU images.
• The DL-CNN generated a likelihood map, which was refined using 3D and 2D level sets for bladder contour generation.
• Performance was evaluated on 173 CTU cases and compared against Haar features and a previous system (CLASS).

Main Results:

• The DL-CNN with level sets achieved superior segmentation accuracy compared to Haar features and the CLASS method.
• Key performance metrics for the DL-CNN method included an 81.9% volume intersection ratio and a 3.6 mm average minimum distance.
• The DL-CNN approach required only a single user input, reducing the manual effort compared to previous methods.

Conclusions:

• Deep learning convolutional neural networks (DL-CNNs) effectively overcome segmentation challenges posed by strong intensity boundaries in CTU.
• The proposed DL-CNN combined with level sets provides a robust and accurate bladder segmentation system.
• This AI-driven approach demonstrates feasibility and improved performance for bladder cancer computer-aided detection.