A Deep Learning Pipeline for Analysis of the 3D Morphology of the Cerebral Small Perforating Arteries from Time-of-Flight 7 Tesla MRI

Le résumé fourni par la source

Abstract The lenticulostriate arteries (LSA) supply important subcortical structures in the brain and are affected in cerebral small vessel disease (CSVD), leading to changes in their morphology. 7 Tesla Time-of-Flight magnetic resonance angiography (7T-TOF-MRA) now allows their visualisation in humans, but current analysis of LSA morphology largely relies on manual tracing on 2D coronal maximum-intensity-projection (MIP) images, which discards significant information from the third spatial dimension. We aimed to develop a semi-automatic pipeline for quantifying the 3D morphology of LSAs from 7T-TOF-MRA in patients with CSVD. We used contrast-enhanced 7T-TOF-MRA data from 15 subjects enrolled in a local CSVD study. Our pipeline consists of two main stages: vessel segmentation and LSA quantification. For segmentation, we fine-tuned a state-of-the-art deep learning model, “DS6”, for vessel segmentation and compared its performance against a classical Frangi filter-based pipeline, Multi-Scale Frangi Diffusive Filter (MSFDF). Both methods were evaluated against manually labelled ground-truth masks in LSA regions. In the LSA quantification stage, the user defines a region-of-interest around LSAs and checks the segmentation. Based on this, the LSA centrelines are extracted, and branch counts, length, tortuosity, and curvature are computed. Additionally, we conducted the traditional LSA analysis using 2D coronal MIPs, and we evaluated the correlation between the results from the 2D and 3D analyses. For vessel segmentation, the fine-tuned DS6 model achieved a mean Dice similarity coefficient (DSC) of 0.814±0.029 during testing, outperforming MSFDF on DSC, sensitivity, and balanced average Hausdorff distance in terms of both mean value and stability. Visual inspection confirmed that DS6 was more sensitive in detecting LSA branches with weak signals. On average, the 15 subjects had 5.9±1.6 LSA stems and 28.7±9.9 branches. The mean length of an LSA branch was 42.5±5.7mm, and mean tortuosity was 1.9±0.2. Finally, the branch counts from 2D and 3D analyses correlated well ( ρ =0.741, p =2.816e-06), whereas the stem count, branch length and tortuosity measurements were significantly different, showing the insufficiency of MIP analysis (stem: ρ =0.230, p =2.207e-01; length: r =0.565, p =1.153e-03; tortuosity: r =0.400, p =2.847e-02). We have developed an open-source semi-automatic pipeline using deep learning for evaluating the 3D morphology of LSAs in CSVD patients from 7T-TOF-MRA. We show that analysing LSA morphology in 3D reveals previously inaccessible aspects of morphology. Our pipeline offers a valuable tool for clinical research studies to characterise the 3D morphology of LSAs in CSVD.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.