An evaluation of variabilities in delineation of organs-at-risk on MR image for MR-only head and neck radiotherapy

Abstract

Background: Radiotherapy (RT) plays an important role in treating head and neck cancers. Though computed tomography (CT) has been the standard imaging modality in RT treatment planning, magnetic resonance (MR) -only planning has come under spotlight due to various potential benefits including lowered radiation dose, elimination of registration error and better alignment of anatomies. However, several significant issues still keep magnetic resonance imaging (MRI) relegated to its secondary role in RT planning. For instance, the lack of accurate electron density information, potential geometric distortions and uncertainties of organs-at-risk (OARs) delineation on MR images. In particular, variability of OAR delineation in head and neck region has not been adequately explored and it remains ambiguous. The aims of this study were to quantify and compare inter-observer variability as well as spatial variation in OAR delineation between and within CT, MR T1W and T2W images.

Method: 15 stereotactic radiosurgery (SRS) cases were used to evaluate contouring between CT and MR images while the MR data in HN region of 10 healthy volunteers were retrieved to compare that on MR T1W and T2W images. Anatomical contouring and image registration were then performed. Using those data sets, we quantified: (1) inter-observer variability within imaging modalities; (2) inter-observer variability between two imaging modalities and; (3) spatial variation in OAR delineation.

Results: 11 and 23 OARs were contoured in cranial and head and neck (HN) region respectively. High occurrence of inter-observer variability was shown in all imaging modalities (100% on CT and 73% on MR in cranial region; 91% on T1W and 70% on T2W images in HN region). MRI achieved comparable or even better agreement than CT for OAR delineation in cranial region. In particular, brainstem and optic chiasm had significantly smaller inter-observer variations in both volumetric and locational aspects on MR images. Inter-observer variabilities between T1W and T2W images were comparable in OAR delineation. Dice similarity coefficient (DSC) values were high for organs with volume exceeding 8cm^3 in cranial region, while organs with volume larger than 5cm^3 attained very satisfactory agreements in HN region.

Conclusion: This study investigated inter-observer variability within and between CT and MRI in cranial region as well as T1W and T2W images in HN region. Further studies on computer-aided automatic segmentation before manual adjustment should be considered. The use of planning organ-at-risk volume (PRV) for smaller organs is advised. T1W and T2W images were comparable in terms of inter-observer variability in HN region. MRI was able to achieve similar or even better agreement than CT for OAR contouring in cranial region. Subsequent researches are essential to extend this result to HN RT.