Florian Thamm
Deep Learning and Image Processing for Stroke Diagnosis with Computed Tomography Angiography
Computed Tomography (CT) is an effective tool, especially in acute situations, which contributes significantly to diagnosing a wide variety of clinical pictures. This includes, with a high prevalence, ischemic stroke, in which an oxygen deficiency is caused by the occlusion of a vessel, e.g., by a thrombus. Typically, an interventional thrombectomy is performed, where the thrombus is surgically removed. Yet, this technically demanding procedure requires precise planning. The vessel pathway must be known, as well as the position of the occlusion. Insight is provided by administering a contrast agent, which makes it possible to highlight blood vessels with high contrast. This type of imaging is called CT angiography (CTA). However, the vascular tree with many branches is complex, and concrete path planning is tedious and time-consuming. In this work, we present methods that use image processing and deep learning to process the CTA data in such a way that (1) the patient’s vascular tree can be interactively explored, (2) vascular occlusions can be automatically and coarsely localized, and (3) the information gain from the available data is increased by enhancing soft tissue contrasts. First, we introduce “VirtualDSA++”, an image processing pipeline for CTA head images used in many ways in the present work. In the first step of the pipeline, seed points are automatically determined for an initial segmentation of blood vessels through region growing. With the help of a vessel atlas, relevant blood vessels are located in the segmentation mask and, for a better overview, marked accordingly in the volume or in the later render view. The next step extends the segmentation to include distal arteries and veins. Subsequently, the vascular mask is skeletonized, and a surface model is computed on which different ways of interaction are possible. In this work, two concrete examples of interactive use were elaborated. First, the shortest paths between two points can be planned and visualized. Second, sub-trees at a certain geodetic distance from a reference point can be interactively hidden from the visualization. The latter makes it possible to restrict the visualization to arteries, which are usually of higher relevance than veins during the diagnosis of strokes. The segmentation of blood vessels, calculated by VirtualDSA++, is also used in the automated detection of large vessel occlusions (LVO). Thrombi lead to a blood flow stop in the affected blood vessel, including contrast agents. In CTA, therefore, affected vessels appear interrupted in their course and, thus, in their segmentation. Using 3-dimensional Convolutional Neural Networks (CNN), an affected vessel tree can hence be classified for the presence of an LVO. However, our experiments have shown that this requires a large number of datasets. Thus, in our initial work, we show that we can significantly increase classification accuracy by aggressively deforming the blood vessel segmentations. Since the deformation is only done on the vessel tree and not in the original scan, deformed trees still represent realistic cases since the vessel trees differ from patient to patient in the actual vessel course anyway. In a follow-up work, we additionally introduced the recombination of vascular trees. The vessel trees are first divided into sub-regions, which are then recombined between patients. With the data obtained, an increase in classification accuracy was again achieved. Dose and reconstruction in CTA scans are designed to highlight vessels. A loss of soft tissue contrast accompanies this. The boundary between the gray and white matter of the brain becomes blurred, allowing only limited views of brain tissue. In our work `SyNCCT”, we present a method that synthesizes non-contrast CT scans from CTA scans while using only one energy level in the CTA scan. The method consists of a GAN-based CNN that, in addition to a segmentation from VirtualDSA++ of the blood vessels to be removed, also receives additional prior knowledge in the form of a statistical estimate of the target image. Quantitatively, the proposed approach was superior to existing methodologies. In a Turing test with physicians, the realism of the images could also be confirmed. In the present work, it was shown that vascular tree extraction enables various applications. Through deep learning, image processing, and their combination, we were able to contribute to modern stroke diagnostics.