Automatic bladder segmentation from CT images using deep CNN and 3D fully connected CRF-RNN.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Accurate bladder segmentation in computed tomography (CT) images is crucial for clinical practice.
• Challenges include significant variations in bladder appearance and low soft-tissue contrast in CT scans.
• Existing automated methods often struggle to achieve high accuracy.

Purpose of the Study:

• To develop a novel deep learning-based approach for accurate automatic bladder segmentation from CT images.
• To introduce a dual-channel preprocessing method to enhance segmentation performance.
• To integrate a convolutional neural network (CNN) with a 3D fully connected conditional random fields recurrent neural network (CRF-RNN).

Main Methods:

• A dual-channel preprocessing technique creating a CT image and an enhanced bladder density map.
• A CNN predicts a coarse voxel-wise bladder score map from the dual-channel image.
• A 3D fully connected CRF-RNN refines the coarse map for a final segmentation result.

Main Results:

• The proposed approach achieved superior segmentation accuracy compared to the state-of-the-art V-net on a clinical dataset.
• Achieved a Dice Similarity Coefficient of 92.24%, outperforming V-net by 8.12%.
• The method produced bladder probability maps with sharper boundaries and more accurate localization.

Conclusions:

• The novel deep learning approach significantly improves bladder segmentation accuracy on clinical CT data.
• Both dual-channel preprocessing and the 3D CRF-RNN integration contribute to the enhanced performance.
• The unified CNN and 3D CRF-RNN network outperforms systems using CRF as a disconnected post-processing step.