Assessment of contour accuracy in head and neck replanning: Deep learning trained model compared with deformable image registration propagation technique

Accurate contouring is crucial for optimal treatment outcomes, whether for nonadaptive radiotherapy with single images or adaptive radiotherapy (ART) with multiple images. For ART there are 2 common approaches for automated segmentation: deformable image registration (DIR) propagation of prior contours from a previous image to a newer replanning image or (ii) deep learning (DL) generated by models trained with datasets. The accuracy of the latter approach is impacted by the size, diversity and quality of the training dataset while the accuracy of the former approach depends on the quality of prior contours, the image contrast between image pairs, and the DIR algorithm used. This study assesses the accuracy of a commercially available pretrained DL model (Mirada DLC04, DLC13, DLC14) and DIR tools (Velocity, MIM, Eclipse) for generating contours in replanning scenarios for adaptive replanning in the head and neck region. Datasets from adaptive replanning in the head and neck region (n = 9 patients) included CTs (n = 18) with clinically approved contours and doses. Manual contour data were compared against deep learning models (Mirada DLC04, DLC13, DLC14) and image registration propagated contours (rigid, deformable with MIM, Velocity, and Eclipse). Evaluation involved (a) contour clinical relevance scores, (b) contour grading scores, (c) assessment of manual and DL contouring style by a Radiation Oncologist Consultant against the Brouwer contouring guideline and (d) accuracy assessment based on geometric and dosimetric metrics, These metrics included dice similarity coefficient (DSC), mean distance to agreement (MDA), Hausdorff distance(HD), volume ratio, and dose ratio. Contours were shortlisted for statistical analysis based on (i) contour relevance (ii) manual contour grading scores and (iii) existing contouring data. Statistical analysis assessed geometric and dosimetric metrics, with the Velocity DIR as the comparator. Contour relevancy scores were highest for spinal cord, parotids, oral cavity, mandible, larynx, and brainstem. Contour grading scores indicated most contours were clinically acceptable contours with minor edits for both manual and DL contours, except for brachial plexus and oral cavity with variation in contouring style described by the Radiation Oncologist. The brainstem and parotid were shortlisted for statistical analysis, with data indicating that: (i) no statistical evidence (all p > 0.1) of dosimetric difference between DL and DIR contours; (ii) geometrically, the DIR algorithm (Velocity) was superior to the DL model (Mirada DLCExpert) in terms of MDA (p = 0.014) and HD (p < 0.001) for parotids, volume difference for brainstem (p = 0.045); (iii) no statistical evidence (all p > 0.1) of geometric or dose difference for parotids and brainstem amongst rigid or deformable registrations. In our study of DL and DIR based contouring for adaptive radiotherapy in the head and neck region, DIR-based contours demonstrated superior geometric accuracy for the parotid glands and brainstem compared to the DL model (Mirada DLCexpert). Among DIR algorithms, no significant differences were observed, except for the MIM DIR volume ratio for brainstem. Our study found no significant dosimetric differences among DIR or DL contouring methods.