Automatisierte Thrombuserkennung anhand dünnschichtig rekonstruierter CCT-Schnittbilder
Automated thrombus detection using thin-slice reconstructed CCT images.
by Hannah Sönne Falk
Date of Examination:2023-07-13
Date of issue:2023-06-26
Advisor:Prof. Dr. Christian Riedel
Referee:Prof. Dr. Thorsten Roland Döppner
Referee:Prof. Dr. Thomas Meyer
Persistent Address:
http://dx.doi.org/10.53846/goediss-9944
Files in this item
Name:HannahSönne_Falk_Dissertation_17.06.2023_eDISS.pdf
Size:3.87Mb
Format:PDF
The following license files are associated with this item:
Abstract
English
In Germany, around 250,000 people suffer a stroke every year. Depending on its severity and extent, a stroke leads to tissue damage and neurological deficits. To evaluate a therapeutic option in the form of intravascular lysis and or mechanical recanalization, multimodality computed tomo- graphy (CT), including native cranial CT, CT angiography of the head and neck vessels, and CT perfusion measurement of the brain parenchyma, is performed. Rapid detection of early signs of infarction and, in particular, viewing the hyperdense arterial sign on native cranial CT as a direct correlate for the vessel obliterating thrombus, is essential for immediate therapeutic decision-making. The hyperdense arterial sign, which is most sensitively found in thin-slice reconstructed CT images, is an indication to perform mechanical recanalization. Individual studies have shown that artificial intelligence methods, particularly forms of deep learning, can be used to detect the hyperdense artery sign. Neural convolutional networks are commonly used for image analysis. Therefore, in this work, a three-dimensional convolutional network was implemented for automated thrombus detection in the anterior cerebral circulation in thin-slice cranial CT datasets. The deep learning library Tensorflow and the interface Keras in the programming language Python were used for the constructed convolutional network. The network structure consisted of an input layer followed by a dropout, two pairs of 3D evolution layers with a maxpoollayer, a batch normalization and a dropout, a 3D evolution layer followed by a flatten layer, a dense layer followed by a dropout and a dense layer at the end. The network was trained on 108 complete cranial CT data sets in which the hyperdense arterial sign was segmented semi-automatically with the MeVisLab program in the previous step. The network was trained on 108 three-dimensional data sets that showed vessel occlusion in the anterior circulation, in the carotid T, M1, or M2 segment. Network performance was tested on 50 data sets with a hyperdense arterial sign in the anterior circulation and on 50 data sets in which no intracranial pathology was present. In the test data set, the neural network marked the hyperdense artery sign very sensitively with a sensitivity of 86%. Analysis of individual parameters showed that the network labels that had the largest volume correlated best with true thrombus. In a further analysis of the incorrect markings, it turned out that the incorrect markings cannot be separated from real thrombi either by their size or by their density values. The neural network made an output in every data set, including a healthy one. In several cases, this output could not be evaluated as meaningful and therefore correctly negative, since it was outside the arterial intracranial vascular system. The limiting factor was that in many cases this marker was located in the intracranial arterial vascular system and can therefore be clinically incorrectly classified and must be considered a false positive finding. In the field of deep learning, there is not just one suitable neural network for a specific question. Depending on the model used, the result for the same question may vary. At the time of implementation and testing of this convolution network, this network was the first to use three-dimensional convolution operations and analyze the entire volume of the skull, thus differing significantly from the literature. The attempt to detect the relatively small hyperdense arterial sign in the entire cranial volume using deep learning methods has not been described in the literature so far. This study is a preliminary study for the automated detection of the hyperdense arterial sign in the entire cranial volume. In the future, the specificity of this network can be increased by further pre-processing of the learning data sets or downstream classification networks.
Keywords: Hyperdense artery sign in stroke; deep learning