A Two-Stage Generative Model with CycleGAN and Joint Diffusion for MRI-based Brain Tumor Detection.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computational Biology

Background:

• Accurate brain tumor detection and segmentation are crucial for diagnosis but challenged by limitations in supervised and unsupervised learning methods.
• Supervised methods demand extensive annotations, while current unsupervised generative models struggle with complete data distribution coverage.

Purpose of the Study:

• To propose a novel unsupervised framework, the Two-Stage Generative Model (TSGM), for enhanced brain tumor detection and segmentation.
• To improve upon existing methods by leveraging generative models and probability distributions for more accurate pathological region identification.

Main Methods:

• The TSGM framework integrates Cycle Generative Adversarial Network (CycleGAN) for generating synthetic abnormal images from healthy ones.
• Variance Exploding using joint probability (VE-JP) is employed to reconstruct healthy images guided by synthetic abnormal data, focusing alterations on pathological areas.
• A thresholding method on image residuals identifies abnormalities, and multimodal results are weighted to boost segmentation accuracy.

Main Results:

• The TSGM framework demonstrated superior segmentation performance across three datasets compared to other unsupervised anomaly detection methods.
• Achieved Dice Similarity Coefficient (DSC) scores of 0.8590 on BraTs2020, 0.6226 on ITCS, and 0.7403 on an In-house dataset.
• The method exhibits strong generalization capabilities in brain tumor segmentation tasks.

Conclusions:

• The proposed TSGM framework effectively improves unsupervised brain tumor detection and segmentation accuracy.
• The integration of CycleGAN and VE-JP offers a robust approach to learning joint probability distributions for conditional generation in medical imaging.
• TSGM shows significant potential for clinical applications requiring precise tumor delineation.