Purpose: Medical image registration is a fundamental and vital task that will affect the efficacy of many downstream clinical tasks. Deep learning (DL)-based deformable image registration methods have been investigated, showing state-of-the-art performance. A test time optimization (TTO) technique was proposed to improve the performance of these DL models. Despite the substantial accuracy improvement with this TTO technique, some regions exhibit large registration errors even after many TTO iterations. This work is to propose a deep interactive registration algorithm for region-specific optimization (RSO).
Methods: We first identified why the TTO technique was slow or failed to improve some regions' registration results. We then proposed a two-level TTO technique, i.e., image-specific optimization (ISO) and region-specific optimization (RSO). The clinician can interactively indicate the region during the registration reviewing process. We further envisioned a three-step DL-based image registration workflow for efficiency and accuracy considerations, i.e., general model-based initial registration, ISO-based rough refinement, and RSO-based further refinement. Phase-resolved CBCT images were acquired from 20 patients, which were then randomly split into 16/4, serving as the training/validation dataset. We manually selected two regions with poor registration results to test the proposed workflow and apply the ISO and RSO methods using 100 and 400 iterations, respectively. We also used the ISO-only method with 500 iterations on the above two regions for comparison. We calculated the root mean squared error (RMSE) to quantify the results.
Results: Visual inspection shows that the warped image associated with the proposed method is closer to the fixed image than the ISO-only method. The quantitative comparison indicates that for the two investigated regions, their RMSEs were reduced from 385/400HU (moving image) to 210/205HU (ISO-only) and further to 131/135HU (proposed).
Conclusion: We experimentally showed that the proposed method outperformed the conventional ISO-only method qualitatively and quantitatively.