Preview

PDF, English (Dissertation Sarina Thomas) - main document Download (35MB) | Terms of use

Abstract

Open reduction and internal fixation is a standard procedure in ankle surgery for treating a fractured fibula. Since fibula fractures are often accompanied by an injury of the syndesmosis complex, it is essential to restore the correct relative pose of the fibula relative to the adjoining tibia for the ligaments to heal. Otherwise, the patient might experience instability of the ankle leading to arthritis and ankle pain and ultimately revision surgery. Incorrect positioning referred to as malreduction of the fibula is assumed to be one of the major causes of unsuccessful ankle surgery. 3D C-arm imaging is the current standard procedure for revealing malreduction of fractures in the operating room. However, intra-operative visual inspection of the reduction result is complicated due to high inter-individual variation of the ankle anatomy and rather based on the subjective experience of the surgeon. A contralateral side comparison with the patient’s uninjured ankle is recommended but has not been integrated into clinical routine due to the high level of radiation exposure it incurs.

This thesis presents the first approach towards a computer-assisted intra-operative contralateral side comparison of the ankle joint. The focus of this thesis was the design, development and validation of a software-based prototype for a fully automatic intra-operative assistance system for orthopedic surgeons. The implementation does not require an additional 3D C-arm scan of the uninjured ankle, thus reducing time consumption and cumulative radiation dose. A 3D statistical shape model (SSM) is used to reconstruct a 3D surface model from three 2D fluoroscopic projections representing the uninjured ankle. To this end, a 3D SSM segmentation is performed on the 3D image of the injured ankle to gain prior knowledge of the ankle. A 3D convolutional neural network (CNN) based initialization method was developed and its outcome was incorporated into the SSM adaption step. Segmentation quality was shown to be improved in terms of accuracy and robustness compared to the pure intensity-based SSM. This allows us to overcome the limitations of the previously proposed methods, namely inaccuracy due to metal artifacts and the lack of device-to-patient orientation of the C-arm. A 2D-CNN is employed to extract semantic knowledge from all fluoroscopic projection images. This step of the pipeline both creates features for the subsequent reconstruction and also helps to pre-initialize the 3D-SSM without user interaction. A 2D-3D multi-bone reconstruction method has been developed which uses distance maps of the 2D features for fast and accurate correspondence optimization and SSM adaption. This is the central and most crucial component of the workflow. This is the first time that a bone reconstruction method has been applied to the complex ankle joint and the first reconstruction method using CNN based segmentations as features. The reconstructed 3D-SSM of the uninjured ankle can be back-projected and visualized in a workflow-oriented manner to procure clear visualization of the region of interest, which is essential for the evaluation of the reduction result. The surgeon can thus directly compare an overlay of the contralateral ankle with the injured ankle.

The developed methods were evaluated individually using data sets acquired during a cadaver study and representative clinical data acquired during fibular reduction. A hierarchical evaluation was designed to assess the inaccuracies of the system on different levels and to identify major sources of error. The overall evaluation performed on eleven challenging clinical datasets acquired for manual contralateral side comparison showed that the system is capable of accurately reconstructing 3D surface models of the uninjured ankle solely using three projection images. A mean Hausdorff distance of 1.72 mm was measured when comparing the reconstruction result to the ground truth segmentation and almost achieved the high required clinical accuracy of 1-2 mm. The overall error of the pipeline was mainly attributed to inaccuracies in the 2D-CNN segmentation. The consistency of these results requires further validation on a larger dataset.

The workflow proposed in this thesis establishes the first approach to enable automatic computer-assisted contralateral side comparison in ankle surgery. The feasibility of the proposed approach was proven on a limited amount of clinical cases and has already yielded good results. The next important step is to alleviate the identified bottlenecks in the approach by providing more training data in order to further improve the accuracy.

In conclusion, the new approach presented gives the chance to guide the surgeon during the reduction process, improve the surgical outcome while avoiding additional radiation exposure and reduce the number of revision surgeries in the long term.