Background: Using flat-detector CT (FD-CT) for stroke imaging has the advantage that both diagnostic imaging and endovascular therapy can be performed directly within the Angio Suite without any patient transfer and time delay. Thus, stroke management could be speeded up significantly, and patient outcome might be improved. But as precondition for using FD-CT as primary imaging modality, a reliable exclusion of intracranial hemorrhage (ICH) has to be possible. This study aimed to investigate whether optimized native FD-CT, using a newly implemented reconstruction algorithm, may reliably detect ICH in stroke patients. Additionally, the potential to identify ischemic changes was evaluated. Methods: Cranial FD-CT scans were obtained in 102 patients presenting with acute ischemic stroke (n = 32), ICH (n = 45) or transient ischemic attack (n = 25). All scans were reconstructed with a newly implemented half-scan cone-beam algorithm. Two experienced neuroradiologists, unaware of clinical findings, evaluated independently the FD-CTs screening for hemorrhage or ischemic signs. The findings were correlated to CT, and rater and inter-rater agreement was assessed. Results: FD-CT demonstrated high sensitivity (95-100%) and specificity (100%) in detecting intracerebral and intraventricular hemorrhage (IVH). Overall, interobserver agreement (κ = 0.92) was almost perfect and rater agreement to CT highly significant (r = 0.81). One infratentorial ICH and 10 or 11 of 22 subarachnoid hemorrhages (SAHs) were missed of whom 7 were perimesencephalic. The sensitivity for detecting acute ischemic signs was poor in blinded readings (0 or 25%, respectively). Conclusions: Optimized FD-CT, using a newly implemented reconstruction algorithm, turned out as a reliable tool for detecting supratentorial ICH and IVH. However, detection of infratentorial ICH and perimesencephalic SAH is limited. The potential of FD-CT in detecting ischemic changes is poor in blinded readings. Thus, plain FD-CT seems insufficient as a standalone modality in acute stroke, but within a multimodal imaging approach primarily using the FD technology, native FD-CT seems capable to exclude reliably supratentorial hemorrhage. Currently, FD-CT imaging seems not yet ready for wide adoption, replacing regular CT, and should be reserved for selected patients. Furthermore, prospective evaluations are necessary to validate this approach in the clinical setting.

Abstract. We present an efficient realization of recent work on unique geodesic paths between tree shapes for the application of matching coronary arteries to a standard model of coronary anatomy in order to label the coronary arteries. Automatically labeled coronary arteries would speed reporting for physicians. The efficiency of the approach and the quality of the results are enhanced using the relative position of detected cardiac structures. We explain how to efficiently compute the geodesic paths between tree shapes using Dijkstra's algorithm and we present a methodology to account for missing side branches during matching. For nearly all labels our approach shows promise compared with recent work and we show results for 8 additional labels.Keywords: coronary labeling, shape space, tree matching. Motivation and OverviewIt is critical that an imaging physician report the anatomical location of pathology in a standard way to the referring physician. A principle goal of automated medical image analysis is the efficient reporting of such findings following established medical guidelines. Criteria have been established for how lesions along the coronary arteries should be reported from CT angiography (CTA) [1,8]. The number of coronary labels varies in different standards but there is agreement on the major labels. The AHA established a standard 15 segment model in 1975 [1]. Our model follows more closely the more recent and more complete models of [8,6]. The physician generally knows which coronary segment contains a lesion but it is time consuming to label images when more than a few labels must be applied. So the goal of automatic coronary labeling is not to inform the physician of the anatomy but to speed the preparation of a report.In order to label the coronaries, our approach will leverage both geometric and topological information to define the correspondence between a labeled model and unlabeled data. We only consider the centerlines of the coronaries and not the coronary lumen. Most prior work labeling vascular or airway trees extracts an abstract graph that captures the topology of the tree but uses limited or no geometric information. A graph matching algorithm is then run to define the best correspondence between nodes in the model and in the unlabeled graph

No abstract

The prediction of the biological activity of a chemical compound is a challenging task in Computational Chemistry and was restricted to vectorial representations of the molecular graph for decades. Kernel functions are positive semidefinite similarity measures that can be defined on arbitrary structured data. This class of similarity functions can be used in kernel-based machine learning algorithms. Interestingly, many graph kernel approaches from Computer Science share properties of traditional similarity measures for chemical compounds, like molecular fingerprints based on paths, cycles and subgraphs.In this work, we present a hybrid technique derived from the Pharmacophore Kernel [1] and Radial Atom Environments [2]. Whereas in the original work [1] atoms were represented by their element number and partial charge, we employ a bounded depth-first-search to enumerate the complete neighbourhood of each non-carbon atom up to specified depth. Therefore, a pharmacophore can be defined by a triangle between three radial fingerprints. This opens the possibility to replace the simple Delta Kernel, which was used for the comparison of the vertices of two pharmacophores in the original work of Mahé et al. In the results section, we benchmark our approach against different state-of-the art graph kernels and the Radial Basis Function Kernel using descriptors calculated with dragonX 1.4 on various QSAR data sets taken from the literature. The models were trained using the machine learning library LIBSVM. The results show that our approach improves the predictive power significantly on many benchmark problems.

No abstract