Bladder cancer segmentation using u-net-based deep-learning

Area of Science:

• Medical Imaging
• Artificial Intelligence in Oncology
• Computational Pathology

Background:

• Accurate segmentation of bladder cancer lesions in CT urography (CTU) is crucial for treatment response assessment.
• Previous methods like deep learning convolutional neural network + level sets (DL-CNN+LS) have limitations.
• Transformer-based models (DATTNet, Med-SAM) are emerging for medical image analysis.

Purpose of the Study:

• To develop and compare novel U-Net based deep learning models for bladder cancer segmentation.
• To evaluate the performance of these models against existing methods (DL-CNN+LS, DATTNet, Med-SAM).
• To simplify the segmentation pipeline by removing the need for level set refinement.

Main Methods:

• Designed and implemented several U-Net based deep learning models for bladder cancer segmentation.
• Proposed a 'Crop U-Net' model using a user-defined box to focus attention on the lesion region.
• Trained and evaluated models using radiologist-annotated 3D contours as the ground truth.
• Compared performance using metrics like Average Jaccard Index (AJI) and Average Minimum Distance (AMD).

Main Results:

• The Crop U-Net model outperformed other investigated models, including DL-CNN+LS, DATTNet, and Med-SAM.
• Crop U-Net achieved an AJI of 48.1±18.0% and AMD of 4.3±3.0 mm on an independent test set.
• This represents a significant improvement over the previous DL-CNN+LS method (AJI 33.2±20.0%, AMD 5.3±2.2 mm).
• The Crop U-Net approach simplified the segmentation pipeline by eliminating the level set refinement stage.

Conclusions:

• The Crop U-Net model demonstrates superior accuracy for bladder cancer segmentation compared to previous methods.
• This AI-driven approach offers a more efficient and accurate tool for bladder cancer treatment response assessment using CTU.
• The simplified pipeline enhances the practicality of using deep learning for clinical decision support in oncology.