Wavelet-Improved Score-Based Generative Model for Medical Imaging.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Signal Processing

Background:

• Score-based generative models (SGMs) excel at inverse problems in medical imaging.
• Training SGMs for medical image reconstruction is challenging due to noisy datasets (e.g., low-dose CT, under-sampled MRI).
• Noise and artifacts in training data degrade SGM performance by affecting gradient estimation.

Purpose of the Study:

• To develop a robust SGM training method for medical image reconstruction using noisy data.
• To enhance the stability and accuracy of SGMs in handling under-determined inverse problems.
• To improve the quality of reconstructed medical images from low-dose CT and under-sampled MRI.

Main Methods:

• Proposed a unified framework integrating a wavelet sub-network with a standard SGM sub-network.
• Implemented a mutual feedback mechanism between the wavelet and SGM sub-networks for accurate score learning from noisy samples.
• Incorporated a regularization constraint during the reconstruction process to further enhance image quality and robustness.

Main Results:

• The proposed wavelet-improved SGM demonstrated superior stability during training with noisy data.
• Achieved significant enhancement in the quality of reconstructed images across various low-dose CT and under-sampled MRI scenarios.
• The method produced results comparable to those trained with clean data, even when using noisy training samples.

Conclusions:

• The integrated wavelet and SGM framework effectively addresses challenges in training SGMs with noisy medical imaging data.
• The proposed method offers a stable and effective solution for high-quality medical image reconstruction, particularly in low-dose and sparse-view applications.
• This approach significantly improves the reliability and precision of reconstructed images, making it valuable for clinical applications.