A Semi-Supervised Multi-Region Segmentation Framework of Bladder Wall and Tumor with Wall-Enhanced Self-Supervised...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Bladder cancer segmentation in MRI is vital for treatment decisions.
• Limited high-quality annotated data hinders data-driven segmentation accuracy.
• Distinguishing bladder wall and tumor is challenging for radiologists.

Purpose of the Study:

• To develop a semi-supervised framework for bladder wall and tumor segmentation using limited annotated and unlabeled MRI data.
• To enhance the discrimination of the bladder wall using self-supervised pre-training.
• To improve the overall segmentation performance for bladder cancer.

Main Methods:

• A semi-supervised multi-region framework incorporating wall-enhanced self-supervised pre-training.
• Utilized limited annotated and unlabeled data for segmentation.
• Introduced contrast consistency and reconstruction observation losses, adaptive learning rates, and post-processing techniques.

Main Results:

• Achieved average Dice Similarity Coefficients (DSCs) of 0.8351 for the bladder wall and 0.9175 for the tumor.
• Demonstrated effective reduction of segmentation artifacts outside the bladder.
• Showcased improved clinical significance of segmentation results.

Conclusions:

• The proposed semi-supervised framework effectively segments bladder walls and tumors in MRI.
• The method leverages limited annotated data and unlabeled data for improved performance.
• Results indicate enhanced clinical utility for bladder cancer diagnosis and treatment planning.