Open Access

Artificial intelligence (AI) is one of the most promising technologies of the post-pandemic era. Cloud computing technology can simplify the process of developing AI applications by offering a variety of services, including ready-to-use tools to train machine learning (ML) algorithms. However, comparing the vast amount of services offered by different providers and selecting a suitable cloud service can be a major challenge for many firms. Also in academia, suitable criteria to evaluate this type of service remain largely unclear. Therefore, the overall aim of this work has been to develop a framework to evaluate cloud-based ML services. We use Design Science Research as our methodology and conduct a hermeneutic literature review, a vendor analysis, as well as, expert interviews. Based on our research, we present a novel framework for the evaluation of cloud-based ML services consisting of six categories and 22 criteria that are operationalized with the help of various metrics. We believe that our results will help organizations by providing specific guidance on how to compare and select service providers from the vast amount of potential suppliers.

We introduce bloomRF as a unified method for approximate membership testing that supports both point- and range-queries. As a first core idea, bloomRF introduces novel prefix hashing to efficiently encode range information in the hash-code of the key itself. As a second key concept, bloomRF proposes novel piecewisemonotone hash-functions that preserve local order and support fast range-lookups with fewer memory accesses. bloomRF has near-optimal space complexity and constant query complexity. Although, bloomRF is designed for integer domains, it supports floating-points, and can serve as a multi-attribute filter. The evaluation in RocksDB and in a standalone library shows that it is more efficient and outperforms existing point-range-filters by up to 4× across a range of settings and distributions, while keeping the false-positive rate low.

Intracranial brain tumors are one of the ten most common malignant cancers and account for substantial morbidity and mortality. The largest histological category of primary brain tumors is the gliomas which occur with an ultimate heterogeneous appearance and can be challenging to discern radiologically from other brain lesions. Neurosurgery is mostly the standard of care for newly diagnosed glioma patients and may be followed by radiation therapy and adjuvant temozolomide chemotherapy. However, brain tumor surgery faces fundamental challenges in achieving maximal tumor removal while avoiding postoperative neurologic deficits. Two of these neurosurgical challenges are presented as follows. First, manual glioma delineation, including its sub-regions, is considered difficult due to its infiltrative nature and the presence of heterogeneous contrast enhancement. Second, the brain deforms its shape, called “brain shift,” in response to surgical manipulation, swelling due to osmotic drugs, and anesthesia, which limits the utility of pre-operative imaging data for guiding the surgery. Image-guided systems provide physicians with invaluable insight into anatomical or pathological targets based on modern imaging modalities such as magnetic resonance imaging (MRI) and Ultrasound (US). The image-guided toolkits are mainly computer-based systems, employing computer vision methods to facilitate the performance of peri-operative surgical procedures. However, surgeons still need to mentally fuse the surgical plan from pre-operative images with real-time information while manipulating the surgical instruments inside the body and monitoring target delivery. Hence, the need for image guidance during neurosurgical procedures has always been a significant concern for physicians. This research aims to develop a novel peri-operative image-guided neurosurgery (IGN) system, namely DeepIGN, that can achieve the expected outcomes of brain tumor surgery, thus maximizing the overall survival rate and minimizing post-operative neurologic morbidity. In the scope of this thesis, novel methods are first proposed for the core parts of the DeepIGN system of brain tumor segmentation in MRI and multimodal pre-operative MRI to the intra-operative US (iUS) image registration using the recent developments in deep learning. Then, the output prediction of the employed deep learning networks is further interpreted and examined by providing human-understandable explainable maps. Finally, open-source packages have been developed and integrated into widely endorsed software, which is responsible for integrating information from tracking systems, image visualization, image fusion, and displaying real-time updates of the instruments relative to the patient domain. The components of DeepIGN have been validated in the laboratory and evaluated in the simulated operating room. For the segmentation module, DeepSeg, a generic decoupled deep learning framework for automatic glioma delineation in brain MRI, achieved an accuracy of 0.84 in terms of the dice coefficient for the gross tumor volume. Performance improvements were observed when employing advancements in deep learning approaches such as 3D convolutions over all slices, region-based training, on-the-fly data augmentation techniques, and ensemble methods. To compensate for brain shift, an automated, fast, and accurate deformable approach, iRegNet, is proposed for registering pre-operative MRI to iUS volumes as part of the multimodal registration module. Extensive experiments have been conducted on two multi-location databases: the BITE and the RESECT. Two expert neurosurgeons conducted additional qualitative validation of this study through overlaying MRI-iUS pairs before and after the deformable registration. Experimental findings show that the proposed iRegNet is fast and achieves state-of-the-art accuracies. Furthermore, the proposed iRegNet can deliver competitive results, even in the case of non-trained images, as proof of its generality and can therefore be valuable in intra-operative neurosurgical guidance. For the explainability module, the NeuroXAI framework is proposed to increase the trust of medical experts in applying AI techniques and deep neural networks. The NeuroXAI includes seven explanation methods providing visualization maps to help make deep learning models transparent. Experimental findings showed that the proposed XAI framework achieves good performance in extracting both local and global contexts in addition to generating explainable saliency maps to help understand the prediction of the deep network. Further, visualization maps are obtained to realize the flow of information in the internal layers of the encoder-decoder network and understand the contribution of MRI modalities in the final prediction. The explainability process could provide medical professionals with additional information about tumor segmentation results and therefore aid in understanding how the deep learning model is capable of processing MRI data successfully. Furthermore, an interactive neurosurgical display has been developed for interventional guidance, which supports the available commercial hardware such as iUS navigation devices and instrument tracking systems. The clinical environment and technical requirements of the integrated multi-modality DeepIGN system were established with the ability to incorporate: (1) pre-operative MRI data and associated 3D volume reconstructions, (2) real-time iUS data, and (3) positional instrument tracking. This system's accuracy was tested using a custom agar phantom model, and its use in a pre-clinical operating room is simulated. The results of the clinical simulation confirmed that system assembly was straightforward, achievable in a clinically acceptable time of 15 min, and performed with a clinically acceptable level of accuracy. In this thesis, a multimodality IGN system has been developed using the recent advances in deep learning to accurately guide neurosurgeons, incorporating pre- and intra-operative patient image data and interventional devices into the surgical procedure. DeepIGN is developed as open-source research software to accelerate research in the field, enable ease of sharing between multiple research groups, and continuous developments by the community. The experimental results hold great promise for applying deep learning models to assist interventional procedures - a crucial step towards improving the surgical treatment of brain tumors and the corresponding long-term post-operative outcomes.

Uncontrolled movement of instruments in laparoscopic surgery can lead to inadvertent tissue damage, particularly when the dissecting or electrosurgical instrument is located outside the field of view of the laparoscopic camera. The incidence and relevance of such events are currently unknown. The present work aims to identify and quantify potentially dangerous situations using the example of laparoscopic cholecystectomy (LC). Twenty-four final year medical students were prompted to each perform four consecutive LC attempts on a well-established box trainer in a surgical training environment following a standardized protocol in a porcine model. The following situation was defined as a critical event (CE): the dissecting instrument was inadvertently located outside the laparoscopic camera’s field of view. Simultaneous activation of the electrosurgical unit was defined as a highly critical event (hCE). Primary endpoint was the incidence of CEs. While performing 96 LCs, 2895 CEs were observed. Of these, 1059 (36.6%) were hCEs. The median number of CEs per LC was 20.5 (range: 1–125; IQR: 33) and the median number of hCEs per LC was 8.0 (range: 0–54, IQR: 10). Mean total operation time was 34.7 min (range: 15.6–62.5 min, IQR: 14.3 min). Our study demonstrates the significance of CEs as a potential risk factor for collateral damage during LC. Further studies are needed to investigate the occurrence of CE in clinical practice, not just for laparoscopic cholecystectomy but also for other procedures. Systematic training of future surgeons as well as technical solutions address this safety issue.

Providing a digital infrastructure, platform technologies foster interfirm collaboration between loosely coupled companies, enabling the formation of ecosystems and building the organizational structure for value co-creation. Despite the known potential, the development of platform ecosystems creates new sources of complexity and uncertainty due to the involvement of various independent actors. For a platform ecosystem to succeed, it is essential that the platform ecosystem participants are aligned, coordinated, and given a common direction. Traditionally, product roadmaps have served these purposes during product development. A systematic mapping study was conducted to better understand how product roadmapping could be used in the dynamic environment of platform ecosystems. One result of the study is that there are hardly any concrete approaches for product roadmapping in platform ecosystems so far. However, many challenges on the topic are described in the literature from different perspectives. Based on the results of the systematic mapping study, a research agenda for product roadmapping in platform ecosystems is derived and presented.