Room: ePoster Forums
Purpose: Due to the increased workload it is important to speed up processes in todayâ€™s clinical routine while not compromising on the quality, e.g. in deformable image registration (DIR). Lately Voxelmorph, a deep learning based DIR for brain MRI datasets was presented. The goal is to investigate if this network can be applied on 3D abdominal CT data via transfer learning, without re-training, and to see if the results are comparable to a conventional DIR algorithm.
Methods: Voxelmorphâ€™s network architecture is publicly available. The latest version was downloaded and implemented on a GPU accelerated machine in Microsoftâ„¢ Azureâ„¢ (IntelÂ® XeonÂ® E5, 112 GB Ram, NvidiaÂ®TeslaÂ® P40). First, the implementation was tested on a publicly available dataset of MRI brain images. Comparable results, when compared to the paper, were achieved. The intensity distributions of five different abdominal CTs of the same patient were then adapted according to the training data since the not-retrained model needs similar distributions to work properly.The resulting deformed CTs were then compared to the results of the DEEDS algorithm. Since segmentations were not available, five statistical metrics: maximum, minimum, standard deviation, mean, and median difference to the fixed image before and after registration, were used to assess the quality of the registration. A negative value shows worse results after registration and vice versa.
Results: The results do not show a clear improvement nor degradation in comparison to the DEEDS algorithm. When averaging over all registrations and all metrics, the difference is reduced by 12.60% using Voxelmorph and by 12.96% with DEEDS.
Conclusion: Based on the statistical metrics applied, Voxelmorph shows similar results compared to DEEDS, while being 40 times faster with 1.7s. Further necessary investigations will include comparisons of the DICE scores and try to further improve the results by retraining the network with abdominal CT data.
Funding Support, Disclosures, and Conflict of Interest: Authors are full-time employees of Varian Medical Systems.