Application of machine learning methods and medical image processing in solving the problem of detecting stenoses of the middle cerebral artery according to computed tomographic angiography data

Abstract

BACKGROUND: Ischemic stroke is a significant contributor to mortality rates in Russia and globally [1]. Computed tomographic angiography is a primary diagnostic tool for ischemic stroke, enabling the identification of stenosis or occlusion in cerebral arteries. The majority of ischemic strokes (51%) occur in the middle cerebral artery region [2], underscoring the growing interest in evaluating blood flow in this area of the brain. The manual detection of stenoses is characterised by subjective evaluation and requires a considerable amount of time. The automation of middle cerebral artery narrowing detection represents a significant challenge in computed tomographic angiography image analysis. AIM: The study aims to develop an algorithm for the automatic detection of stenoses in the middle cerebral artery on DICOM images of computed tomographic angiography based on the application of artificial neural networks, vascularity assessment and skeletonization algorithms. MATERIALS AND METHODS: A total of 262 computed tomographic angiography series from patients at the N.V. Sklifosovsky Emergency Medical Research Institute were analyzed. Of these, 94 series exhibited stenosis in the M1/M2 segment of the middle cerebral artery. The image processing was conducted using an artificial neural network with a CFPNet-M architecture [3]. The reconstruction of the vascular tree was based on the calculation of the "vesselness" measure [4] with subsequent skeletonization of the identified structures. RESULTS: In the initial stage, a neural network for the segmentation of the middle cerebral artery basin was trained. The training array was generated using the MNI152 template with affine transformations and subsequent expert evaluation. In this case, the IoU (Intersection over Union) measure was 0.81. The primary objective was the segmentation of the middle cerebral artery vascular tree, which was achieved through the use of the vesselness filter, followed by an evaluation of voxel intensities and the identification of the connected object with the longest length. The next stage involved the construction of the skeleton of the middle cerebral artery. This entailed determining the centerline of the vessel and representing the resulting skeleton as a graph with the vessels as edges and their bifurcation points as vertices. The subsequent stage was the calculation of morphological features (diameter, area, and perimeter) in the cross-sectional plane for each segment (the area between the bifurcation points). Finally, the area of constriction was determined based on the analysis of the behavior of the segment cross-sections and the identification of any deviation from the threshold value. The overall accuracy of the algorithm was 79.39% (95% confidence interval 73.98–84.12), with a sensitivity of 80.85% (95% confidence interval 71.44–88.24) and a specificity of 78.57% (95% confidence interval 71.59–84.52). CONCLUSIONS: Thus, we developed an algorithm for the detection of stenoses in the M1/M2 segment based on the segmentation of the middle cerebral artery basin, the assessment of vesselness, and the skeletonization of the vascular tree. The application of the developed algorithm in practice, after its validation and clinical approval, will simplify the routine evaluation of computed tomographic angiography images by radiologists and provide an opportunity to obtain an objective assessment of the stenosis area.