Program Information
Development of a Neural Network Based Quality Evaluator for MR-CT Fusion of Brain Images
J Wu*, K Mund , G Yan , B Lu , University of Florida, Gainesville, FL
Presentations
WE-RAM3-GePD-I-3 (Wednesday, August 2, 2017) 10:30 AM - 11:00 AM Room: Imaging ePoster Lounge
Purpose: To develop a neural network based registration quality evaluator (RQE) for MR-CT fusion of brain images for patients with brain tumors receiving external beam radiation therapy.The RQE can be used to identify “bad” solutions and increase the robustness of an automated fusion method.
Methods: The MR and CT images of 10 brain tumor patients were used. The RQE was pattern classifier based on a two-layer feed-forward neural network. A supervised training was used to train the classifier based on pre-calculated registration solutions and geometric features of sampled cost function. The registration solutions were compared to their “golden” solutions to determine their registration errors. A solution in the training dataset was determined as a “good” or a “bad” one based on a user-defined error threshold. For each patient, 500 registration attempts were performed between the MR and CT images by optimizing their mutual information. Those registrations differ only by randomly generated initial transformation parameters within a reasonable clinical range. RQE performance was evaluated using test data with registration results that were not used in training. The impacts of the sampling step size and the user-defined error threshold on the performance of RQE were also investigated.
Results: The RQE scored well, with the average value for sensitivity and specificity of 0.943 (range 0.807-1.000) and 0.903 (range 0.641-1.000), respectively, when the threshold for the mean-voxel-deviation was 20 mm and the sampling step size was 10 mm.
Conclusion: The proposed RQE can potentially be used to improve the robustness of the current 3D image fusion systems by rejecting “bad” registration solutions. The RQE is not patient specific and only need to be constructed once for a specific anatomical site. Incorporating the RQE into an existing automated image fusion system would only incur a small increase in the overall registration time.
Contact Email: