Synthetic data generation with Worley-Perlin diffusion for robust subarachnoid hemorrhage detection in imbalanced CT...

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Subarachnoid hemorrhage (SAH) detection in CT scans is critical for patient outcomes.
• Imbalanced datasets pose a significant challenge for developing accurate SAH detection models.
• Existing generative models for SAH sample synthesis suffer from limitations like overfitting and noise distortion.

Purpose of the Study:

• To propose a novel generative model for producing high-quality SAH samples to enhance SAH CT detection performance.
• To address the limitations of previous methods in handling imbalanced datasets and generating realistic SAH images.

Main Methods:

• Introduction of the Worley-Perlin Diffusion Model (WPDM) utilizing Worley-Perlin noise for SAH image synthesis.
• WPDM is designed to overcome the homogeneity of Gaussian noise and distortion of Simplex noise.
• A faster variant, WPDM_Fast, is developed to optimize generation speed without sacrificing image quality.

Main Results:

• WPDM demonstrated significant improvements in classification accuracy across datasets with varying imbalance ratios.
• A classifier trained with WPDM-generated samples achieved an F1-score of 0.857 on a 1:36 imbalance ratio.
• This performance surpassed the state-of-the-art by 2.3 percentage points.

Conclusions:

• WPDM effectively generates high-quality, realistic SAH images, overcoming limitations of previous noise-based models.
• The model significantly enhances classification performance in imbalanced scenarios.
• WPDM offers a robust solution for improving SAH CT detection in clinical practice.