Session Item

The MR-linac adapts treatment plans to accommodate inter-fraction anatomical changes. Patients will experience further intra-fraction motion which will impact the final delivered dose. We aim to assess the performance of a hybrid intensity- and structure-based DIR algorithm for inter- and intra-fraction prostate MR images using geometric and dosimetric assessments.

We generated manual contours of the bladder, rectum, and CTV on 3D T2 MR images of 25 patients with prostate cancer treated with 30Gy/5 SBRT. In addition to the reference MR scan, three MR scans were obtained for each patient per fraction: session start (adapt), end of plan adaptation (verify), and during beam delivery (beam-on). We contoured 400 MR images. We recomputed adapt doses on the verify and beam-on MR scans. We performed DIRs for each patient between the reference MR and all fraction images, as well as within fraction adapt-to-verify (A-to-V) and adapt-to -beam-on (A-to-B). We use a hybrid DIR algorithm with two strategies: image intensity alone, and image intensity with controlling structures. The latter uses Chamfer matching to optimize the registration of the bladder, rectum, and CTV ROIs. For a subset of patients, we included controlling point-of-interest (POIs) in the DIR. We automatically generate the POIs based on an optimized mesh representation of the bladder adapted across all MR images. We compare manual and mapped contours using max distance to agreement (mDTA), dice similarity coefficient (DSC), and difference of DVHs (dDVH).

In Fig. 1 we compare DSC between manual and mapped structures for the A-to-V and A-to-B DIRs using intensity-only and combined intensity-structure strategies. Use of the controlling ROIs improved the DSC and reduced variance. For the A-to-B intensity-only DIR, mDTA was 0.34±0.13 cm (CTV), 2.1±1.2 cm (bladder), and 0.9±0.6 cm (rectum). For the intensity-structure DIR, the values were 0.14±0.06 cm, 0.25±0.13 cm, and 0.3±0.6 cm, respectively. In Fig. 2 we present the dDVH between the manual and A-to-B mapped contours. For 14 patients, use of the controlling ROIs notably reduces dose differences across the entire dDVH. For 9 patients, the combined intensity-structure DIR resulted in worse ROI alignment (mDTA of 0.22±0.11 cm, 2.6±2.8 cm, 0.8±0.7 cm for A-to-B). By using controlling POIs, we were able to improve these DIRs (mDTA of 0.20±0.09 cm, 0.27±0.14 cm, 0.7±0.7 cm). The DIR between the reference and beam-on images using the intensity-only DIR produced a DSC of 0.89±0.04, 0.77±0.16, 0.83±0.07 and a difference in dose metrics of 59±125 cGy (CTV D98), 2±115 cGy (bladder D5cc), and -38±114 cGy (rectum D1cc).

We demonstrate that the hybrid DIR algorithm can incur errors in contour propagation which will impact application to dose warping and accumulation. Controlling ROIs and POIs improve the DIR and consequentially require ROI segmentation on registered images. This motivates implementation of robust automated contouring for dose accumulation work.