Implementing a deep learning model for automatic tongue tumour segmentation in ex-vivo 3-dimensional ultrasound volumes

Three-dimensional (3D) ultrasound can assess the margins of resected tongue carcinoma during surgery. Manual segmentation (MS) is time-consuming, labour-intensive, and subject to operator variability. This study aims to investigate use of a 3D deep learning model for fast intraoperative segmentation of tongue carcinoma in 3D ultrasound volumes. Additionally, it investigates the clinical effect of automatic segmentation. A 3D No New U-Net (nnUNet) was trained on 113 manually annotated ultrasound volumes of resected tongue carcinoma. The model was implemented on a mobile workstation and clinically validated on 16 prospectively included tongue carcinoma patients. Different prediction settings were investigated. Automatic segmentations with multiple islands were adjusted by selecting the best-representing island. The final margin status (FMS) based on automatic, semi-automatic, and manual segmentation was computed and compared with the histopathological margin. The standard 3D nnUNet resulted in the best-performing automatic segmentation with a mean (SD) Dice volumetric score of 0.65 (0.30), Dice surface score of 0.73 (0.26), average surface distance of 0.44 (0.61) mm, Hausdorff distance of 6.65 (8.84) mm, and prediction time of 8 seconds. FMS based on automatic segmentation had a low correlation with histopathology (r = 0.12, p = 0.67); MS resulted in a moderate but insignificant correlation with histopathology (r = 0.4, p = 0.12, n = 16). Implementing the 3D nnUNet yielded fast, automatic segmentation of tongue carcinoma in 3D ultrasound volumes. Correlation between FMS and histopathology obtained from these segmentations was lower than the moderate correlation between MS and histopathology.