Search

Computer-assisted surgery (CAS) is a novel treatment modality that allows clinicians to create personalized treatment plans and virtually design implants, surgical guides, and radiotherapy boluses. However, even though CAS is a powerful means to optimize surgery, it currently requires a lot of tedious and time-consuming manual work by experienced medical engineers to ensure the quality of the CAS workflow. This thesis therefore focuses on developing artificial intelligence algorithms, specifically convolutional neural networks (CNNs), to automate two main image processing tasks required in maxillofacial CAS: CT image segmentation and computed tomography (CT) artifact correction. Chapter 2 describes a CNN approach to segment bony structures in 20 different CT scans. All CT were acquired from patients that had previously undergone craniotomy, thus posing a significant challenge for the CNN to correctly recognize the pathological shape of the bones. Even though this segmentation tasks was validated on a challenging dataset, only two regions of interest (i.e., bone or background) were segmented. In order to demonstrate the ability of CNN to take more regions into account, a CNN was developed in Chapter 3 to segment cone-beam computed tomography (CBCT) scans into the jaw, the teeth and background. To date, many different training strategies have been proposed in literature to train CNNs on medical images. However, it remains unclear which strategy yields the best segmentation performances. To elucidate on this topic, eight different CNN training strategies were evaluated and compared in Chapter 4. The CNNs described in this chapter were trained to segment anatomical structures in simulated and experimental cone-beam CT scans. These experiments demonstrated that the best strategy is generally to train three separate CNN and combine their segmentation results through majority voting. In Chapter 5, an approach is described to deal with strong metal artifacts in CBCT scans during the image segmentation step required during CAS. In particular, a mixed-scale dense convolutional neural network (MS-D network) was implemented to segment the bony structures of the mandible and the maxilla. The MS-D network described in this chapter clearly outperformed a widely-used clinical segmentation method (i.e., snake evolution), and produced comparable results to alternative deep learning benchmarks. A novel symmetry-aware deep learning approach is proposed in Chapter 6 to reduce high cone-angle artifacts in CB CT images. In this approach, a CNN was trained using radial cone-beam CT slices to exploit the symmetry of high cone-angle artifacts. This allowed training a CNN to reduce the complex 3D cone-angle artifacts using only 2D slices as input. In this chapter, it is demonstrated that this symmetry-aware dimensionality reduction improves the performance and robustness of CNNs when reducing high cone-angle artifacts in cone-beam CT scans. Deep learning, and specifically CNNs, have found remarkable successes in various mage processing tasks. The goal of Chapter 7 was to review all studies that have been published in which neural network approaches were developed for CT image reconstruction, bone segmentation, and surgical planning. Although various neural network approaches were identified, the majority of studies (66%) applied CNNs. Interestingly, all of these CNNs were published from 2016 onwards, indicating the rapid paradigm shift this field has undergone. Nevertheless, much research is still required to make deep learning an integral part of the CAS workflow. In conclusion, this thesis contributes to enhance the understanding of CNN approaches for medical image processing. Nevertheless, many interesting challenges and questions remain to incorporate CNN’s as an integral part of the maxillofacial CAS routine. I therefore hope that this thesis will inspire fellow researchers to take on these exciting challenges.