ESTRO 2024 - Abstract Book

S3040

Physics - Autosegmentation

ESTRO 2024

1 RMIT University, School of Science, Melbourne, Australia. 2 Peter MacCallum Cancer Centre, Physical Sciences Department, Melbourne, Australia. 3 University of Melbourne, Sir Peter MacCallum Department of Oncology, Melbourne, Australia

Purpose/Objective:

There is limited literature on individual vertebra growth in paediatric patients following radiotherapy, largely because measuring the growth of individual vertebrae is time-consuming [1]. Even fewer studies discuss normative paediatric growth models. This study is designed to create and investigate an automated technique for detecting landmarks on vertebrae bodies to assist with growth measurement in the paediatric population. The study's primary objective is to compare two methods for detecting growth measurement landmarks on paediatric patient CT images.

Material/Methods:

Initial vertebrae segmentation was performed using a nnUnet-based algorithm. For its training, 150 cases were sourced from the VerSe dataset with ground-truth vertebral segmentations, and an additional 70 paediatric cases were obtained from the TCIA paediatric segmentation dataset [2-4]. Ground truth masks for the TCIA cases were generated through a semi-automated approach in 3D Slicer and subsequently underwent manual verification. To enhance performance on paediatric CT data, both adult and paediatric datasets were integrated during the training phase. Once segmentation was completed, landmarks on each vertebra, including the anterior superior and inferior corners, posterior superior and inferior corners, and midpoints on the superior and inferior end plates, were extracted from the centroid sagittal slice of each vertebra. The use of each vertebral centroid slice avoids the influence of scoliosis or misalignment of the spine with the mid-sagittal plane. Two methodologies were adopted for this extraction. In the first, masks derived from auto-segmentation were analysed, augmented by a python-based image processing script that identified mask corner points and the midpoints on the boundary of the mask. In the second, a U-Net-based AI auto-detection model was developed for landmark identification. Manual curation within 3D Slicer was employed to generate its training set which consisted of sagittal CT slices and csv files containing the landmark coordinates. This AI landmark detection algorithm was trained using 120 CT sagittal images. The performance of the analytical script based method was compared to the AI approach using Euclidean distances between the ground truth landmarks and detected landmarks. To assess and compare accuracy, landmarks were detected on 10 test cases, unseen by the AI, encompassing both lumbar and thoracic vertebrae. Figure 1 shows an example of a CT image with vertebral segmentation and classification. An individual vertebra mask and its 6 analytically detected landmarks are shown. Alongside is a complete spine landmark set with AI-selected points in blue overlayed with ground truth points in red. As illustrated in Figure 2 which shows data for 1020 landmark detections, the U-Net-based method generally outperformed the analytical approach. The spread of errors was typically broader for the analytical method with larger magnitude outliers, even when the mean errors were not statistically significantly different (11/17 vertebrae, 2-tailed test, p>0.05). The poorer performance of the image analysis script was attributed to its sensitivity to the quality and smoothness of the bone segmentation contours generated by the nnUnet auto-segmenter. Some outliers were related to imperfect separation of the vertebral processes and the vertebral body in the mask. Furthermore, it Results: