TL;DR
This paper introduces a novel deep learning approach to automatically estimate optimal parameters for an iterative image reconstruction algorithm (PDEM) in computed tomography (CT). By training a neural network to predict these parameters from projection data, the method achieves improved image quality and reduces the need for manual tuning, making CT reconstruction more robust and efficient.
What Problem Does This Paper Address?
Computed tomography (CT) reconstruction, particularly iterative methods, relies on objective functions with parameters that significantly impact image quality. The PDEM algorithm, an extension of MLEM, uses two parameters that need to be carefully tuned based on the noise level of the projection data and the characteristics of the object being imaged. Manually finding these optimal parameters is time-consuming and requires expertise. This paper addresses the challenge of automating this parameter estimation process to improve the efficiency and robustness of CT image reconstruction.
Methods
The core of this work is a novel neural network architecture that acts as a "learned optimizer." This architecture consists of two main components:
- Optimizer (Λ): A neural network that takes the projection data ($y$) as input and outputs the optimal parameters ($\lambda = (\gamma, \alpha)$) for the PDEM algorithm. This network is designed with a series of dense layers, batch normalization, and dropout layers to stabilize learning and prevent overfitting.
- Optimizee (Ω): This component is the PDEM iterative image reconstruction algorithm itself. It takes the projection data ($y$) and the estimated parameters ($\lambda$) from the optimizer to reconstruct the image. Notably, the optimizee does not contain trainable parameters; it's essentially the fixed reconstruction process.
The entire system ($\Theta$) is trained end-to-end using a deep learning approach. The network learns to predict parameters that minimize a loss function (squared L2 norm) comparing the reconstructed image to a ground truth image. A key innovation in the optimizee's structure is the use of multiplicative connections within its "residual blocks," which the authors suggest may help mitigate the vanishing gradient problem.
Datasets
The study utilizes simulated data for training and evaluation.
- Training and Validation Data: 5000 pairs of input projection data and corresponding ground truth images were generated. The ground truth images were based on the Shepp-Logan phantom, with various attributes (intensity, size, position, rotation) deformed to create diverse phantoms. Projection data was simulated using a Radon transform model with 90 directions and 95 detectors, with signal-to-noise ratios (SNRs) uniformly distributed between 20 dB and 50 dB.
- Test Data: 5 test data pairs were used, with varying SNRs and phantom shapes, to evaluate the learned optimizer's performance.
- Larger Image/Sinogram Data: For demonstrating scalability, projections were generated for larger images (128x128) with 180 directions, using magnified versions of the initial phantoms.
Key Insights and Ideas
- Automated Parameter Tuning: The most significant insight is that a neural network can effectively learn to estimate optimal parameters for iterative CT reconstruction algorithms directly from projection data, eliminating the need for manual tuning.
- Learned Optimization Architecture: The proposed architecture, separating parameter estimation (optimizer) from the reconstruction process (optimizee), offers a more straightforward and potentially more efficient approach compared to traditional recurrent neural network-based learned optimization methods.
- Multiplicative Connections: The use of multiplicative connections within the optimizee's residual blocks is a novel design choice that may contribute to better gradient flow during training, especially for deep iterative processes.
- Scalability to Larger Problems: The study demonstrates that parameters learned on smaller datasets can be effectively applied to reconstruct larger images, suggesting a practical pathway to reduce training costs for real-world applications.
- Parameter Dependence: The research confirms that optimal parameters are influenced by both the noise level of the projection data and the shape of the object being reconstructed, highlighting the complexity of manual tuning.
Key Results
- The proposed neural network successfully learned to estimate parameters for the PDEM algorithm, leading to significantly improved image reconstruction quality compared to the standard MLEM algorithm (which uses fixed parameters $\gamma=1, \alpha=1$).
- In cases with high noise (low SNR), the learned parameters resulted in a substantial decrease in the evaluation function (L2 norm difference between reconstructed and true image) over iterations, outperforming MLEM.
- For test cases where MLEM performed relatively well, the proposed method still achieved better or comparable results, demonstrating its robustness.
- The learned optimizer, trained on smaller projections, was successfully applied to reconstruct larger images, showing a significant reduction in granular artifacts and improved contour reproducibility compared to MLEM.
Limitations and Future Work
- Optimizer Complexity: The authors acknowledge that designing a more efficient optimizer A is a future challenge, as the current dense layer requires a large number of parameters, leading to high computational costs for training.
- Parameter Upper Bounds: While the learned parameters can temporarily exceed stable calculation tolerances, the authors implemented an upper limit to avoid numerical instability. Further investigation into handling such extreme values might be beneficial.
- Practical Application of Larger Images: The paper demonstrates the feasibility of using parameters learned from smaller projections for larger ones. However, developing an appropriate method to appropriately reduce the size of large original projections for training remains a practical challenge for actual use.
- Generalization to Real-World Data: The current study relies on simulated data. Future work could explore the application and validation of this method on real clinical CT data.
Summary updated