Session Item

In recent years, there has been an increasing interest in the application of magnetic resonance (MR) in radiation therapy (RT), essentially due to the benefits that using magnetic resonance imaging (MRI) provides. MRI allows an optimal soft tissue differentiation of target tumours and organs-at-risk (OARs). Moreover, MRI-only RT is desirable when repeated scanning can be helpful during treatment to monitor early response and evaluate possible OARs and target changes, i.e., for replanning during treatment while reducing patient X-ray exposures. Accurate OARs delineation is of critical importance to achieve the most efficient radiation planning process, mainly to ensure adequate dose coverage whilst minimising dose to OARs. However, manual delineation of each structure is a time-consuming effort. Moreover, this segmentation is subject to observer bias and intra/interobserver variability. While that is still the most widely used approach, software-based auto-segmentation techniques are being increasingly used clinically. This work aims to evaluate the use of ADMIRE® software (research version 3.28, Elekta AB, Sweden) for the multi-atlas-based segmentation of H&N structures on MR images.

To create a multi-atlas database, the MRI scans of eleven subjects were acquired using a T1w 3D VIBE Dixon gradient echo sequence on a 1.5T Magnetom Aera scanner (Siemens Healthcare, Germany). The MRI sequences were acquired in the same setup of CT simulation, including a thermoplastic mask fixed on a flat table. Twenty-six structures (OARs, bone and air) were contoured on each MRI dataset with the support of a radiation oncologist and a neuroradiologist. The geometric accuracy of three different algorithms implemented on ADMIRE® (STAPLE, Patch Fusion and Random Forest) was evaluated in terms of Dice similarity coefficient (DSC), using a leave-one-out cross-validation approach and considering the OAR manually segmented as the gold standard.

The mean DSC (± 1 std. dev.) of the different methods were reported in Tab. 1. The Random Forest (RF) algorithm showed better results compared to the other two and proved to be a reliable tool for automatic delineation. Nevertheless, the RF algorithm has not proven capable of correctly contouring small structures (lens, cochlea and optical nerves). However, the time required for contouring these structures is significantly short and has almost no impact on the total contouring time. An example auto-contouring result of the Random Forest algorithm is reported in Figure 1.

This study demonstrated the feasibility of accurate atlas-based automatic OARs segmentation that could lead to significant time-saving in the overall MRI-only RT workflow. Therefore, improvements in the automatic segmentation of these OARs are required to reduce the time of the manual editing process, and further investigations are necessary to improve its performance for small structures.