Deep learning adapted acceleration for limited-view photoacoustic image reconstruction
Optics Letters
|April 1, 2022
Summary
This study introduces a novel deep learning approach to improve photoacoustic (PA) computed tomography image quality from limited views. The method accelerates reconstruction and enhances image fidelity, achieving superior results with fewer iterations.
Area of Science:
- Medical Imaging
- Biomedical Engineering
- Computational Imaging
Background:
- Limited-view data in photoacoustic (PA) computed tomography leads to low-quality images due to geometric constraints.
- Model-based methods with regularization are employed to address these reconstruction challenges.
Purpose of the Study:
- To develop a fast and high-quality reconstruction method for limited-view PA data.
- To combine mathematical variational models with deep learning for accelerated and regularized PA image reconstruction.
Main Methods:
- A novel model-based method integrating variational models and deep learning was proposed.
- A deep neural network was designed to optimize gradient descent steps for data consistency.
- The unrolled reconstruction procedure was accelerated and regularized.
Main Results:
- The proposed method achieved superior performance, with a Structural Similarity Index (SSIM) improvement of over 0.05 compared to other model-based methods.
- High-quality PA images were obtained with significantly fewer iterations (three steps).
- In vivo results demonstrated a high SSIM value (0.94), indicating excellent image fidelity and robustness.
Conclusions:
- The combined variational and deep learning approach effectively addresses the limited-view issue in PA CT.
- The method offers accelerated reconstruction and robust, high-quality image generation for PA imaging applications.
