Elsevier

Radiotherapy and Oncology

Volume 158, May 2021, Pages 112-117
Radiotherapy and Oncology

Guidelines
An evaluation of MR based deep learning auto-contouring for planning head and neck radiotherapy

https://doi.org/10.1016/j.radonc.2021.02.018Get rights and content

Highlights

  • MR based deep learning auto-contouring is effective for head and neck OAR delineation.

  • A model created on diagnostic MR scans works well on similar scans and on RTP scans but needs optimisation on MR Linac sequences.

  • Performance of the MR based model is superior to the CT based model on respective radiotherapy planning scans.

Abstract

Introduction

Auto contouring models help consistently define volumes and reduce clinical workload. This study aimed to evaluate the cross acquisition of a Magnetic Resonance (MR) deep learning auto contouring model for organ at risk (OAR) delineation in head and neck radiotherapy.

Methods

Two auto contouring models were evaluated using deep learning contouring expert (DLCExpert) for OAR delineation: a CT model (modelCT) and an MR model (modelMRI). Models were trained to generate auto contours for the bilateral parotid glands and submandibular glands. Auto-contours for modelMRI were trained on diagnostic images and tested on 10 diagnostic, 10 MR radiotherapy planning (RTP), eight MR-Linac (MRL) scans and, by modelCT, on 10 CT planning scans. Goodness of fit scores, dice similarity coefficient (DSC) and distance to agreement (DTA) were calculated for comparison.

Results

ModelMRI contours improved the mean DSC and DTA compared with manual contours for the bilateral parotid glands and submandibular glands on the diagnostic and RTP MRs compared with the MRL sequence. There were statistically significant differences seen for modelMRI compared to modelCT for the left parotid (mean DTA 2.3 v 2.8 mm), right parotid (mean DTA 1.9 v 2.7 mm), left submandibular gland (mean DTA 2.2 v 2.4 mm) and right submandibular gland (mean DTA 1.6 v 3.2 mm).

Conclusion

A deep learning MR auto-contouring model shows promise for OAR auto-contouring with statistically improved performance vs a CT based model. Performance is affected by the method of MR acquisition and further work is needed to improve its use with MRL images.

Section snippets

Methods

Bilateral parotid glands and submandibular glands (SMGs) were contoured by the same clinician on one hundred anonymised oropharynx T2 weighted (T2W) 2 dimensional (2D) diagnostic MR scans. Diagnostic T2W MR scans were used due to the ease of obtaining sufficient numbers to investigate the model’s performance. To ensure adequate visualisation of structures, oropharynx scans were excluded if the images contained artefacts or there was tumour involvement of the parotid or SMGs. Contours were peer

Comparison of manual contours and modelMRI

ModelMRI automated contours showed good agreement with manual contours defined on T2W 2D diagnostic and T2W Dixon 2D radiotherapy planning scans for bilateral parotid glands and SMGs (mean DSC ≥ 0.80). The agreement was lower for the T2W 3D MRL for the bilateral parotid glands (mean DSC 0.70). There was a lack of overlap between modelMRI automated and manual contours on the MRL images for the left and right SMG (mean DSC of 0.10 and 0) as shown in Supplementary Table 2. The overlap of manual

Comparison of modelMRI and modelCT

ModelMRI contours showed better agreement with manual contours on MR radiotherapy planning scans compared with modelCT contours on CT radiotherapy planning scans. An improvement in mean DTA and reduction in SD for bilateral parotid and SMGs was observed except for the left SMG which was similar between the two modalities (see Fig. 4). The mean DSC values to compare the model and manual contours were similar for the bilateral parotid and left SMGs irrespective of the model used (mean DSC > 0.7).

Discussion

This study is the first to evaluate the cross-acquisition of a deep learning MR auto-contouring model on multiple MR sequences for head and neck OAR auto contouring. The work shows that MR based deep learning auto-contouring models show promise as an aid to clinician OAR contouring but are sensitive to the MR sequence used. ModelMRI performed well when applied to T2W 2D diagnostic and radiotherapy planning scans, with a mean DSC score of 0.80, and DTA of ≤ 2 mm.

The MR auto contouring model was

Conclusion

MR based deep learning auto-contouring models show promise as an aid to clinician OAR contouring. A model trained on diagnostic MR images has been shown to work well on RTP images as well as diagnostic images. However extending this to MRL images shows that these models remain sensitive to the MR sequence used. Further work is needed to optimise the MRL image acquisitions and evaluate the model in a larger cohort prior to evaluating its use in adaptive re-planning.

Declaration of Competing Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

Acknowledgements

The Christie NHS Foundation Trust is supported by a Cancer Research UK Centres Network Accelerator Award Grant (A21993) to the ART-NET consortium. Professor West and Professor van Herk are supported by the NIHR Manchester Biomedical Research Centre. This work was also supported by Cancer Research UK via funding to the Cancer Research Manchester Centre [C147/A25254].

The MR-Linac images were taken as part of the PRIMER imaging protocol funded by RMH/ICR NIHR biomedical research centre.

References (24)

Cited by (0)

View full text