Localization and Standard Plane Regression of Vertebral Bodies in Intra-Operative CBCT Volumes

Spinal fractures account for 4.3% of all bone fractures and occur in 30 – 50% of people with an age above 50 [1, 2]. These fractures often remain unnoticed and do not induce issues for the affected. However, this behaviour poses a direct risk. Untreated damages increase the probability and severity of future fractures which lead to pain, reduced quality of life or increased mortality [1, 3]. Moreover, about 10% of the vertebral fractures are directly linked to injuries of the spinal cord [2]. Spinal cord damages result in a wide range of disorders including movement restrictions, loss of sensitivity, autoimmune diseases or even total paralysis. Their treatment highly depends on the phase and the advance of the damage. Therefore, an early damage detection improves the chances of a positive outcome [4]. Computed Tomography (CT) represents the gold standard for diagnosis of bone fractures. But especially for smaller clinics, the utilization within the intraoperative environment is denied due to high costs, a restricted access to the patient as well as substantial space requirements [5]. To allow image acquisition without the pitfalls of CT acquisition in the intraoperative suite, mobile CTs have become the clinical standard. These systems acquire 2D projections during a 190°rotation around the patient, which are used for reconstruction of a 3D volume [6].

An essential task in image-guided surgery is the generation of the so-called standard planes. These standard planes are used to obtain a standardized view of the anatomical structure showing its key features [7]. This helps to facilitate the evaluation process as well as reduces the risk to overlook damages. The standard plane regression is done manually by the reader. Although this action is used as a normalization procedure, it is dependent on the physician and leaves room for mistakes [8]. Furthermore, the physician can not adjust the image output and perform the surgery on the patient at the same time. Therefore, the adjustment needs to be done first and the surgery can only start afterwards. This increases the overall surgery duration. To fasten this process, standardize it and make it less physician dependent, an automation of the vertebral body detection as well as standard plane regression for cone beam computed tomography (CBCT) volumes is needed.

To the best of our knowledge there are no publications on this exact topic yet. However, there are many possible building blocks for an approach to automate standard plane regression. One possibility is to use segmentation for that task. Shi et al. proposed a two-step algorithm to localize and segment vertebral bodies in CT images using a combination of a 2D and a 3D U-net [9]. Thomas et al. proposed an assistance system forankle evaluation. A sliding window approach in combination with two consecutive 3D U-Nets is used on CBCT volumes to segment the ankles and regress the standard planes for each one [7]. Another idea is to determine the standard plane parameters based on a preceding bounding box prediction of the respective object. For that purpose, Jaeger et al. introduced a Retina U-Net. The fusion of a Retina Net one-stage detector with the U-Net architecture, is used to predict bounding boxes of lesions in CT lung images [10]. In real-time object detection the gold standard is the You Only Look Once (YOLO) algorithm. The YOLO processes the whole input in one step. Multiple layers are used to predict each object’s surrounding bounding box as well as their class affiliation [11].

This thesis aims to design a framework for the localization of vertebral bodies followed by a standard plane regression in intra-operative CBCT volumes based on deep learning algorithms. The detection as well as the determination should be fast and accurate at the same time. Therefore, existing algorithms are utilized and compared against one another. The U-Net architecture, the Retina U-Net idea as well as the YOLO algorithm will be analyzed to realize the task at hand. In detail, the thesis will comprise the following work items:

[1] Ghada Ballane et al. Worldwide prevalence and incidence of osteoporotic vertebral fractures. OsteoporosisInternational, 28(5):1531–1542, 2017.

[2] Zhao Wen Zong et al. Chinese expert consensus on the treatment of modern combat-related spinal injuries. Military Medical Research, 6(1), 2019.

[3] Neil Binkley et al. Lateral vertebral assessment: A valuable technique to detect clinically significant vertebral fractures.Osteoporosis International, 16(12):1513–1518, 2005.

[4] Katari Venkatesh et al. Spinal cord injury: pathophysiology, treatment strategies, associated challenges, and future implications.Cell and Tissue Research, 377(2):125–151, 2019.

[5] Stefan Wirth et al. C-arm-based mobile computed tomography: a comparison with established imaging ont he basis of simulated treatments of talus neck fractures in a cadaveric study.Computer Aided Surgery,9(1-2):27–38, 2004.

[6] Jan Von Recum et al. Die intraoperative 3D-C-Bogen-Anwendung.Unfallchirurg, 115(3):196–201, 2012.

[7] Sarina Thomas et al. Computer-assisted contralateral side comparison of the ankle joint using flat panel technology. Technical report, 2021.

[8] Lisa Kausch et al. Toward automatic C-arm positioning for standard projections in orthopedic surgery. International Journal of Computer Assisted Radiology and Surgery, 15(7):1095–1105, 2020.

[9] Dejun Shi et al. Automatic Localization and Segmentation of Vertebral Bodies in 3D CT Volumes with Deep Learning.ISICDM 2018: Proceedings of the 2nd International Symposium on Image Computing andDigital Medicine, pages 42–46, 2018.

[10] Paul Jaeger et al. Retina U-Net: Embarrassingly Simple Exploitation of Segmentation Supervision for Medical Object Detection. Technical report, 2018.

[11] Joseph Redmon et al. You Only Look Once: Unified, Real-Time Object Detection. InProceedings of theIEEE conference on computer vision and pattern recognition, pages 779–788, 2016.