Accuracy of automatic structure propagation for daily magnetic resonance image-guided head and neck radiotherapy

Abstract

Abstract Background and purpose Deformable image registration (DIR) and contour propagation are used in daily online adaptation for hybrid MRI linac (MRL) treatments. The accuracy of the propagated contours may vary depending on the chosen workflow (WF), affecting the amount of required manual corrections. This study investigated the impact of three different WFs of contour propagations produced by a clinical treatment planning system for a high-field MRL on head and neck cancer patients. Methods Seventeen patients referred for curative radiotherapy for oropharyngeal cancer underwent standard CT-based dose planning and MR scans in the treatment position for planning (pMR), and at the 10th (MR10), 20th (MR20) and 30th (MR30) fraction (±2). The primary tumour, a metastatic lymph node and 8 organs at risk were manually delineated on each set of T2 weighted images. Delineations were repeated one month later on the pMR by the same observer to determine the intra-observer variation (IOV). Three WFs were used to deform images in the treatment planning system for the high-field MRL: In WF1, only the planning image and contours were used as a reference for DIR and propagation to MR10,20,30. The most recently acquired image set prior to the daily images was deformed and uncorrected (WF2) versus manually corrected (WF3) structures propagated to the session image. Dice similarity coefficient (DSC), mean surface distance (MSD) and Hausdorff distance (HD) were calculated for each structure in each model. Results Population median DSC, MSD and HD for WF1 and WF3 were similar and slightly better than for WF2. WF3 provided higher accuracy than WF1 for structures that are likely to shrink. All DIR workflows were less accurate than the IOV. Conclusions WF1 and WF3 provide higher accuracy in structure propagation than WF2. Manual revision and correction of propagated structures are required for all evaluated workflows.