Abstract: WPDM overcomes the limitations of Gaussian and Simplex noise-based models, generating high-quality, realistic SAH images. It significantly enhances classification performance in imbalanced settings, providing a robust solution for SAH CT detection.
Introduction
In medical imaging, "data imbalance" presents a significant challenge for deep learning models. Conditions such as subarachnoid hemorrhage (SAH)—a critical, life-threatening stroke—are statistically rarer than normal scans. This scarcity often leads AI models to develop a bias toward "healthy" results, potentially resulting in fatal misdiagnoses. While traditional data augmentation and generative tools like Denoising Diffusion Probabilistic Models (DDPMs) exist, they often struggle with medical textures; standard Gaussian noise is too uniform, and Simplex noise can cause anatomical distortions.
Methodology: The Worley–Perlin Diffusion Model (WPDM)
The researchers introduced the Worley–Perlin Diffusion Model (WPDM) to address these limitations. The core innovation is the replacement of standard white noise with a specialized blend of Perlin noise (organic gradients) and Worley noise (cellular patterns). This combination better mimics the complex, non-uniform textures found in human brain tissue.
The WPDM Pipeline
- Preprocessing: A U-Net model performs "skull stripping" with 99.13% Dice-score accuracy, allowing the generative model to focus exclusively on brain matter.
- Noise Generation: The model blends Perlin and Worley noise to create a more anatomically relevant noise profile for the diffusion process.
- WPDMFast: To ensure clinical practicality, the authors optimized noise generation using 2D grid computations and GPU acceleration, achieving an 18x speedup (from 3.67s to 0.02s per image pair).
- Downstream Classification: Synthetic SAH images are used to rebalance training sets to a 1:1 ratio, which are then used to train a DenseNet-121 classifier.
Experimental Results and Performance
The study evaluated the model using CT scans from the National Institute of Informatics, Japan. To test robustness, the authors created five training sets with increasing imbalance ratios: 1:7, 1:12, 1:18, 1:28, and 1:36.
Key Findings
- Superior Accuracy: At the most extreme imbalance ratio (1:36), the WPDM-augmented classifier achieved an F1-score of 0.857, outperforming the previous state-of-the-art (ImbSAM) by 2.3 percentage points.
- Image Realism: The Fréchet Inception Distance (FID) for WPDM was 19.350, significantly lower (better) than standard DDPM (24.625) and AnoDDPM (26.137), indicating higher synthetic image quality.
- Structural Integrity: Unlike Simplex-based models, the Worley-Perlin blend successfully introduced realistic hemorrhage textures without compromising the underlying brain morphology.
Discussion and Future Work
The "aha moment" of this research is that the geometry of the noise used to degrade an image during training dictates the model's ability to reconstruct realistic features. By using non-uniform noise, the model learns more complex structural relationships.
While WPDM provides a clear advantage in imbalanced scenarios, future work will focus on improving ultra-fine detail resolution and decoupling specific brain shapes from hemorrhage patterns to further enhance the indistinguishability of synthetic lesions from real pathology.
Summary updated