Purpose: The alignment of the region of interest is of primary concern in radiotherapy. Although quantitative assessment can be obtained by manually contouring the same structures on both fixed and moving images, this process is time-consuming and observer-dependent. Here, we established an efficient and stable evaluation tool using deep learning to evaluate the accuracy of image registration quantitatively and objectively.
Methods: Ten patients with head and neck tumors were enrolled. Two sets of CT images for each patient were acquired in different time, and then were registered with rigid and deformable methods, respectively. Eight OARs were segmented by oncologists and the deep learning method in each image, named as Manual-fixed, Manual-moving, Auto-fixed and Auto-moving respectively. Then, both Manual-moving and Auto-moving were introduced to the transformation matrices to get Manual-moving-trans and Auto-moving-trans. The similarities between Manual-fixed and Manual-moving-trans were quantified by the geometric index, including Dice similarity coefficient (DSC) and mean distance to agreement (MDA), which was set as the ground truth (GT) in this study. The same analysis was applied to Auto-fixed and Auto-moving-trans, obtaining the value of DSC and MDA. Finally, it was compared with the GT to verify its reliability. The paired t-test was performed for statistical analysis. A p-value less than 0.05 was statistically significant.
Results: For the rigid registration, ΔDSC for all OARs were smaller than 0.012%, and all ΔMDA less than 0.2mm. There were no significant differences between GT and evaluation tool. In the deformable case, ΔDSC for all OARs were within 0.02%, and ΔMDA were always less than 0.09mm. There were no significant except for the spinal cord and the mandible.
Conclusion: The proposed deep learning-based evaluation tool could substitute the conventional evaluation method with the quantitative measurement, reflecting the accuracy of registration stably and efficiently in the interested region.