Deep-learning convolutional neural network: Inner and outer bladder wall segmentation in CT urography.

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Radiology

Background:

• Accurate bladder wall segmentation is crucial for CT urography (CTU) analysis.
• Existing methods may face challenges with inconsistent bladder wall distinctions.

Purpose of the Study:

• To develop and evaluate a computerized segmentation tool for inner and outer bladder walls using CTU data.
• To integrate a deep learning convolutional neural network (DL-CNN) into an image analysis pipeline.

Main Methods:

• Retrospective analysis of 172 CTU cases, split into training and testing sets.
• Training a DL-CNN on regions of interest to identify bladder wall pixels.
• Applying DL-CNN likelihood maps within a cascaded level sets method for segmentation.
• Comparing DL-CNN-assisted segmentation against 3D manual outlines.

Main Results:

• The DL-CNN-assisted level sets method demonstrated effective segmentation of inner and outer bladder walls.
• Achieved average volume intersection of 86.9% for the inner wall and 87.5% for the outer wall in the test set.
• Outer wall segmentation showed improvement compared to previous methods.

Conclusions:

• DL-CNN-assisted level sets provide an effective tool for bladder wall segmentation in CTU.
• Potential for over-segmentation and incorrect segmentation of adjacent structures (e.g., prostate) exists.
• The tool shows promise for improving bladder wall analysis in CT urography.