Deep Learning Based Modality-Independent Intracranial Aneurysm Detection

Abstract

Early detection of intracranial aneurysms (IAs) allows early treatment and therefore a better outcome for the patient. Deep learning-based models trained and executed on angiographic scans can highlight possible IA locations, which could increase visual detection sensitivity and substantially reduce the assessment time. Thus far methods were mostly trained and tested on single modality, while their reported performances within and across modalities seems insufficient for clinical application. This paper presents a modality-independent method for detection of IAs on MRAs and CTAs. First, the vascular surface meshes were automatically extracted from the CTA and MRA angiograms, using nnUnet approach. For IA detection purpose, the extracted surfaces were randomly parcellated into local patches and then a translation, rotation and scale invariant classifier based on deep neural network (DNN) was trained. Test stage proceeded by mimicking the surface extraction and parcellation, and the results across parcels were aggregated into IA detection heatmap of the entire vascular surface. Using 200 MRAs and 300 CTAs we trained and tested three models, two in cross modality setting (training on MRAs/CTAs and testing on CTAs/MRAs, respectively), while the third was a mixed-modality model, trained and tested on both modalities. The best model resulted in a 96% sensitivity at 0.81 false positive detections per image. Experimental results show that proposed approach not only significantly improved detection sensitivity and specificity compared to state-of-the-art methods, but is also modality agnostic, may aggregate information across modalities and thus seems better suited for clinical application.