Development of proton-density magnetic resonance angiography using deep-learning models as new cerebral angiography for the evaluation of intracranial aneurysms: Preliminary study

Introduction: In our previous study, we demonstrated that high-resolution proton-density (HR-PD) outperformed time-of-flight magnetic resonance angiography (TOF-MRA) in differentiating intracranial infundibula from aneurysms. HR-PD is also non-invasive imaging and has the excellent vascular visualization under the superior signal-to-noise and spatial resolution. However, unlike TOF-MRA, the high-resolution proton-density images remains limited in enhancement capabilities due to a lack of reconstructed images such as maximum intensity projection (MIP) and vascular volume rendering (VR). To address this, we propose proton-density magnetic resonance angiography (PD-MRA) having VR and source images as new angiography using an efficient labeling method, a deep learning (DL)-based vascular segmentation model.
Methods: A total of 25 paired HR-PD and TOF-MRA scans were collected under IRB approval. All scans were acquired using a 3T MRI scanner (Ingenia CX, Philips Healthcare) with the following scanning parameters: FOV 180×180 mm²; reconstructed voxel size 0.25×0.25×0.25 mm³. HR-PD vascular labels were generated through the following steps: 1) Affine registration of TOF-MRA to HR-PD. 2) Intensity inversion of HR-PD image and followed by addition to the TOF-MRA image. 3) Vessel map extraction using a region-growing algorithm. 4) Manual removal of venous regions and noisy pixels in the vessel map using an annotation tool. The dataset was split into 20 training cases and 5 test cases. For segmentation, we employed the 3D nnU-Net model. Input volumes were divided into patches of size (192, 192, 64) and normalized using z-score normalization. The PD-MRA which is results from the DL models and the manual segmentation were compared using Dice similarity coefficients, and 5 cases were reviewed on PD-MRA and compared with TOF-MRA in terms of differentiation of aneurysms from infundibula.
Results: The Dice similarity coefficient for the model performance achieved a high score of 86.86% on the test set. All of 5 test cases showed precise differentiation between aneurysms and infundibula on PDMRA, whereas it appeared poorly defined on TOF-MRA.
Discussion & Conclusion: We present PD-MRA as a new cerebral angiography with VR and source images using a DL based method. It successfully segmented vessels and aneurysms, while effectively capturing small branching vessels critical for distinguishing infundibula even though it is preliminary research. Clinically, the approach may aid radiologists in diagnosing aneurysm/infundibulum more accurately.