Intracranial steno-occlusive lesion detection on magnetic resonance angiography images

Intracranial steno-occlusive disease has received attention in pathology, imaging and many other aspects in recent years because of its risk.With the assistance of deep learning, exploring the detection of intracranial steno-occlusive lesion is valuable in clinical diagnosis.However, the narrow character of intracranial steno-occlusive lesion renders the classical detection methods not applicable for this task.We hope to propose a method to enhance the detection effect on non-invasive time-of-flight magnetic resonance angiography.For each axial single-layer slice from time-of-flight magnetic resonance angiography, the operation of multi-scale division is applied to address the narrowness of lesions.The images contain different ranges are detected by the models of corresponding scales separately, and the multi-scale detection results are then fused into the final results on slices.Our in-house dataset contains 200 patients from 6 different hospitals in this study, and the qualitative and quantitative results are evaluated on the test set of it.Under the premise of enhanced precision value, the recall value has significant improvement up to more than fifty percent compared with other competing methods.The recall values achieve 0.855 for slice-based detection results and 0.929 for case-based detection results separately.Overall, we propose a multi-scale detection method to address the narrow size of intracranial steno-occlusive lesion.The experimental results show the feasibility of our method and multi-scale detection through the fusion from different scale ranges of the images can enhance the detection effect of intracranial steno-occlusive lesion.