Room: Karl Dean Ballroom B1
Purpose: Iterative CT reconstruction is often formulated as an optimization problem with an objective function containing a data-fidelity term and a regularization term to ensure image quality under clinically favorable conditions, such as dose reduction. There are always parameters to control trade-offs between different terms. Tuning these parameters is a critical, yet unsolved problem. Manual parameter tuning is not only tedious, but become impractical with many parameters involved. Motivated by recent advances in deep learning-based decision making, we propose a deep reinforcement learning (DRL) approach to automatically adjust parameters with human-level intelligence.
Methods: We consider an example problem of CT reconstruction with pixel-wise-total-variation regularization. Each pixel has one parameter to adjust. A parameter-tuning policy network (PTPN) is established to map an image patch to actions specifying direction and amplitude of parameter-tuning for its center pixel. We define a reward function that favors actions improving image quality using ground truth as reference. PTPN is trained using simulated data via end-to-end DRL to learn parameter-tuning strategy maximizing reward over the large action space. When PTPN is applied, it repeatedly observes each image patch and determines parameter adjustment, similar to a human. The process iterates until image quality cannot be further improved. We test PTPN using independent simulation and experimental cases.
Results: In all cases, parameters are initialized with random values. In simulation studies, PTPN intelligently adjusts parameters, yielding ~3% lower root-mean-square error (RMSE) than initial parameters. Compared to manually tuned parameters, PTPN-guided reconstruction is slightly better (~0.5% lower RMSE). Similar behavior is observed in experimental cases. Moreover, parameters tuned by PTPN appeared similar to the optimal parameters derived theoretically, demonstrating the effectiveness of PTPN.
Conclusion: DRL-based PTPN can tune parameters of iterative CT reconstruction in a human-like manner. The resulting image is similar to or better than those under manual parameter tuning.
Not Applicable / None Entered.