Machine Learning-Based Tumor Contour Propagation for MRI Adaptive Radiotherapy of Gliobastoma

Abstract

The adoption of adaptive strategies for glioblastoma radiotherapy based on frequent MR imaging is challenged by considerable workflow and planning requirements. Promising techniques such as those based on small margin conformal treatments, dose escalation, or anatomic adaptation may not be reliably implemented without efficient processes to mitigate the overall workload. This study investigated a machine learning based algorithm to propagate gross tumor volume (GTV) contours defined at planning to subsequent imaging timepoints. Twenty glioblastoma patients participated in this prospective, research ethics board approved study. All patients had undergone surgical tumor resection and completed four MRIs: at planning (Fx0), fractions 10 and 20 of radiotherapy (Fx10 and Fx20), and one month post-radiotherapy (P1M) using a T1-weighted sequence with Gadolinium contrast. On each image, the primary GTV was manually contoured as the surgical resection cavity plus any residual enhancing regions. The manual contours at Fx10, Fx20, and P1M were compared to GTVs propagated using three different methods: 1) Six degree of freedom rigid registration of the Fx0 timepoint (RIGID). 2) Local cross-correlation deformable registration of the Fx0 timepoint (DEF). 3) A machine learning method using an atlas developed with the images and GTVs of all previous timepoints (ML). For all methods, additional disease foci that developed at Fx10, Fx20, or P1M were excluded from the analysis. The GTV generated by each of the propagation methods was compared to the manual contour using Dice similarity coefficient (DSC) and Hausdorff distance metrics. Rigid based propagation poorly captured the tumor dynamics during radiotherapy. Across all patients, the median DSC and Hausdorff distance at (Fx10, Fx20, P1M) was (0.79, 0.71, 0.67) and (8.84, 9.81, 11.21) mm, respectively, for the RIGID method. The DEF method increased this overall agreement to (0.88, 0.85, 0.82) and (7.56, 9.51, 11.21) mm. Finally, the ML method improved these results to (0.88, 0.90, 0.88) and (7.56, 7.56, 6.94) mm. Among outlier patients, the ML technique was also superior; the minimum DSC at (Fx10, Fx20, P1M) was (0.59, 0.58, 0.54), (0.70, 0.61, 0.62), and (0.70, 0.69, 0.65) for the RIGID, DEF, and ML methods, respectively. Machine learning based tumor contour propagation is a promising and efficacious technique to facilitate MRI based adaptive glioblastoma radiotherapy. Especially as the frequency of soft tissue imaging increases, the accuracy of machine learning based atlas approaches will improve. Tools and algorithms such as that presented here will be necessary to best maximize the benefit of adaptive strategies based on frequent imaging.