ESTRO 2024 - Abstract Book

S1380

Clinical - Head & neck

ESTRO 2024

The inter-observer delineation was not significantly different between both observers (P>0.05) measured using DSC, volume difference, and Hausdorff distance. The overall delineations from both observers were larger in the DL scan compared to the conventional scan (averaged volume=13.3cc (observer 1) and 12.6cc (observer 2) in DL scans; averaged volume=12.53 cc (observer 1) and 12.0 cc (observer 2) in the conventional scans) (P>0.05). This could likely be due to the larger slice thickness used in the DL scan. Slightly higher DSC was also observed in the delineation of parotid gland (both left and right) (DSC=0.82) in the DL scan compared to the conventional 3D-VIBE scan (DSC=0.78) (P>0.17). Similar Hausdorff distance was also found across different OARs delineation between both scans (average Hausdorff=1.48, 95% Hausdorff=4.43 for DL; average Hausdorff=1.55, 95% Hausdorff=4.68) (P>0.12). There was no significant difference in inter-observer delineation variability among different OARs delineation (P>0.47). Despite a more visually-noticeable motion artifact and a relatively shorter acquisition time in DL scans compared to conventional scans, the results show comparable inter-observer delineation between both DL scans and conventional scans. The delineation of the DL could potentially be further improved with the use of immobilization, hence reducing the motion artifact. This study has limitations, including a small sample size, differences in spatial resolution, and the lack of radiotherapy immobilization during the scan. Additionally, the delineation similarity of the deep-learning-based 2D TSE images was compared to conventional 3D VIBE images rather than comparing it to the 2D TSE image without applying the deep-learning-based super-resolution algorithm. Further studies with the same sequence will be conducted using a larger sample size and including radiotherapy immobilization and setup.

Conclusion:

Based on the results of this pilot study, the deep learning-based super-resolution post-processing algorithm enabled faster MR acquisition with higher spatial resolution and still resulted in comparable delineation variability to the conventional images.

Keywords: MR-guided radiation therapy, deep learning, MRI

References:

[1]Corradini, S., Alongi, F., Andratschke, N. et al. MR-guidance in clinical reality: current treatment challenges and future perspectives. Radiat Oncol 14, 92 (2019). https://doi.org/10.1186/s13014-019-1308-y

[2] Chaudhari AS, Fang Z, Kogan F, Wood J, Stevens KJ, Gibbons EK, Lee JH, Gold GE, Hargreaves BA. Super-resolution musculoskeletal MRI using deep learning. Magn Reson Med. 2018 Nov;80(5):2139-2154. doi: 10.1002/mrm.27178. Epub 2018 Mar 26.

2164

Digital Poster

A trial of low-dose 5-Fu and cisplatin and PD-1 in metastatic nasopharyngeal carcinoma (NCT04890522)