Optimising a 3D convolutional neural network for head and neck computed tomography segmentation with limited training data.

Background and purpose: Convolutional neural networks (CNNs) are increasingly used to automate segmentation for radiotherapy planning, where accurate segmentation of organs-at-risk (OARs) is crucial. Training CNNs often requires large amounts of data. However, large, high quality datasets are scarce. The aim of this study was to develop a CNN capable of accurate head and neck (HN) 3D auto-segmentation of planning CT scans using a small training dataset (34 CTs). Materials and Method: Elements of our custom CNN architecture were varied to optimise segmentation performance. We tested and evaluated the impact of: using multiple contrast channels for the CT scan input at specific soft tissue and bony anatomy windows, resize vs. transpose convolutions, and loss functions based on overlap metrics and cross-entropy in different combinations. Model segmentation performance was compared with the inter-observer deviation of two doctors' gold standard segmentations using the 95th percentile Hausdorff distance and mean distance-to-agreement (mDTA). The best performing configuration was further validated on a popular public dataset to compare with state-of-the-art (SOTA) auto-segmentation methods. Results: Our best performing CNN configuration was competitive with current SOTA methods when evaluated on the public dataset with mDTA of ( 0.81 ± 0.31 ) mm for the brainstem, ( 0.20 ± 0.08 ) mm for the mandible, ( 0.77 ± 0.14 ) mm for the left parotid and ( 0.81 ± 0.28 ) mm for the right parotid. Conclusions: Through careful tuning and customisation we trained a 3D CNN with a small dataset to produce segmentations of HN OARs with an accuracy that is comparable with inter-clinician deviations. Our proposed model performed competitively with current SOTA methods.

Use of a novel atlas for muscles of mastication to reduce inter observer variability in head and neck radiotherapy contouring.

PURPOSE/OBJECTIVE(S): Trismus is caused by injury to the masticatory muscles resulting from cancer or its treatment. Contouring these muscles to reduce dose and radiation related trismus can be problematic due to interobserver variability. This study aimed to evaluate the reduction in interobserver variability achievable with a new contouring atlas. MATERIALS/METHODS: The atlas included: medial and lateral pterygoids (MP, LP), masseter (M) and temporalis (T) muscles, and the temporo-mandibular joint (TMJ). Seven clinicians delineated five paired structures on CT scans from 5 patients without the atlas. After ≥5 weeks, contouring was repeated using the atlas. Using contours generated by the clinicians on the same 5 CT scans as reference, dice similarity coefficient (DSC), mean distance-to-agreement (DTA) and centre of mass (COM) difference were compared with and without the atlas. Comparison was also performed split by training grade. Mean and standard deviation (SD) values were measured. RESULTS: The atlas reduced interobserver variability for all structures. Mean DTA significantly improved for MP (p = 0.01), M (p < 0.01), T (p < 0.01) and TMJ (p < 0.01). Mean DTA improved using the atlas for the trainees across all muscles, with the largest reduction in variability observed for the T (4.3 ±â€¯7.1 v 1.2 ±â€¯0.4 mm, p = 0.06) and TMJ (2.1 ±â€¯0.7 v 0.8 ±â€¯0.3 mm, p < 0.01). Distance between the COM and interobserver variability reduced in all directions for MP and T. CONCLUSION: A new atlas for contouring masticatory muscles during radiotherapy planning for head and neck cancer reduces interobserver variability and could be used as an educational tool.

Setup uncertainties of anatomical sub-regions in head-and-neck cancer patients after offline CBCT guidance.

PURPOSE: To quantify local geometrical uncertainties in anatomical sub-regions during radiotherapy for head-and-neck cancer patients. METHODS AND MATERIALS: Local setup accuracy was analyzed for 38 patients, who had received intensity-modulated radiotherapy and were regularly scanned during treatment with cone beam computed tomography (CBCT) for offline patient setup correction. In addition to the clinically used large region of interest (ROI), we defined eight ROIs in the planning CT that contained rigid bony structures: the mandible, larynx, jugular notch, occiput bone, vertebrae C1-C3, C3-C5, and C5-C7, and the vertebrae caudal of C7. By local rigid registration to successive CBCT scans, the local setup accuracy of each ROI was determined and compared with the overall setup error assessed with the large ROI. Deformations were distinguished from rigid body movements by expressing movement relative to a reference ROI (vertebrae C1-C3). RESULTS: The offline patient setup correction protocol using the large ROI resulted in residual systematic errors (1 SD) within 1.2 mm and random errors within 1.5 mm for each direction. Local setup errors were larger, ranging from 1.1 to 3.4 mm (systematic) and 1.3 to 2.5 mm (random). Systematic deformations ranged from 0.4 mm near the reference C1-C3 to 3.8 mm for the larynx. Random deformations ranged from 0.5 to 3.6 mm. CONCLUSION: Head-and-neck cancer patients show considerable local setup variations, exceeding residual global patient setup uncertainty in an offline correction protocol. Current planning target volume margins may be inadequate to account for these uncertainties. We propose registration of multiple ROIs to drive correction protocols and adaptive radiotherapy to reduce the impact of local setup variations.