br The drawback of these newer
The drawback of these newer techniques however, is that delin-eation is time consuming due to the complex head and neck anat-omy. Delineation of OARs alone can take up to one hour, and of TVs up to two hours , which will be more for unexperienced radia-tion oncologists (ROs). On top of that, correct delineation is essen-tial for optimal treatment planning and is one of the weakest links in the chain of actions needed to treat a patient with RT. Firstly, it ASP-1517 is mainly performed manually and therefore susceptible to intra- and
inter-observer variability (IOV) . Secondly, variations in OAR delineation may influence the treatment plan, including the dose to OARs, which can also impact results of multicentre trials [11–13]. Thirdly, delineation errors remain present during the entire RT course so their impact can be larger than expected. IOV in TV and OAR delineation affect quality of RT, treatment outcomes and evaluation of clinical research [14,15]. The introduction of delineation guidelines for OARs has improved IOV , although Brouwer et al.  showed that there was still room for improvement.
In previous research, automated segmentation using machine learning approaches has been widely investigated to overcome drawbacks of manual segmentation procedures in medical imaging . Available algorithms in current RT software are mainly atlas-based methods, which incorporate prior knowledge in the form of atlases and are registered to the daily images using deformable image registration (DIR). In particular for HNC patients, atlas-based models achieved acceptable results for segmentation of OARs [18,19], but encountered difficulties with patient variability specifically in the tumour regions, due to a fixed number of atlases. Significant editing of contours is required, which does not improve the segmentation workflow [20,21]. Recently, deep learning approaches based on convolutional neural networks are gaining popularity thanks to their successes in many segmentation tasks in medical imaging [22,23], including in RT [20,24–27]. Ibragimov et al. was the first to propose a tri-planar patch-based convolu-tional neural network for segmentation of OAR for HNC and proved to achieve state-of-the-art results over atlas-based models. Nikolov et al. used a 3D UNET and proved the generalisability of the net-work to an independent test set while Zhu et al. expanded a 3D UNET for whole volume segmentations of head and neck anatomy, eliminating the need for patch-based approaches. Although this previous research investigated different deep learning methods and their performance on OAR segmentation, they did not evaluate the clinical benefits of using the network in daily clinical practice.
Precise delineation of OARs in HNC is necessary for accurate radiotherapy treatment planning, correct interpretation of dose volume histograms (DVH) and reduction of therapeutic variability. The aim of this study was to evaluate the potential of a 3D convo-lutional neural network (CNN) for automated delineation of OARs most commonly delineated in HNC patients, which could signifi-cantly reduce delineation time and the burden of human interven-tion and IOV. The clinical implementation of the validated automated delineation tools could eventually result in a shorter interval between simulation and start of RT, affect treatment capacity and facilitate paradigm shifts such as online adaptive planning.
Tumour site and TNM staging for the 15 HNC patients in the study.
Patients were recruited between August and November 2018 and included consecutive patients with a newly diagnosed HNC, scheduled for RT and without total laryngectomy. In total, 15 patients were included in the study (see Table 1 for patient charac-teristics). Each patient underwent a contrast-enhanced planning CT scan in the supine position with custom thermoplastic mask for immobilization, according to the conventional clinical protocol. The CT images were made on a multidetector-row spiral CT scan-ner (Somatom Sensation Open, 40 slice configuration; Siemens Medical Solutions, Erlangen, Germany). The acquisition parameters were: 120 kvP/230 mAs (quality reference mAs with CARE Dose4D), no gantry tilt, spiral mode, rotation time 1 s, 40 detector rows at 0.6 mm intervals, table speed 21.6 mm/rotation (pitch = 0.9), reconstruction interval 3 mm using Kernel B30s med-ium smooth, matrix size 512 512, pixel spacing 0.97 0.97 mm.
3D convolutional neural network
Using international delineation consensus guidelines [28,29], a CNN was previously developed and trained for automated delin-eation of 16 OARs in contrast-enhanced planning CT images of HNC patients based on a training set of 70 cases  (see Appendix A for more details). This CNN, was applied to all 15 images in this study to assess the clinical benefits. We refer to the original, unmodified delineations generated by the CNN as ‘‘automated delineations” further on. Note that the CNN was trained on con-trast enhanced CT scans from the same scanner and same institu-tion as the scans for our clinical evaluation.