Author: Gayoung Kim, Junghoon Lee 👨🔬
Affiliation: Department of Radiation Oncology and Molecular Radiation Sciences, Johns Hopkins University 🌍
Purpose: 3D time-of-flight magnetic resonance angiography (TOF-MRA) is widely used for visualizing cerebrovascular structures. Accurate segmentation of cerebrovascular structures is critical for reliable diagnosis and treatment. However, manual segmentation is challenging and time-consuming due to the intricate topology and fine details of the vascular structure. This study aims to address these challenges by developing a deep learning based method for cerebrovascular segmentation.
Methods: The proposed cerebrovascular segmentation model integrates convolutional module and multi-head self-attention based transformer module in the encoder. These modules separately extract 3D spatial and contextual features of the vascular structure from TOF-MRA images, which are then merged at each layer. Skip connections between the encoder and decoder are utilized to refine the prediction map. To account for diversity in TOF-MRA images, we utilized the publicly available CerebrOvascular SegmenTAtion (COSTA) dataset, which includes six sub-datasets collected from different centers.
Results: The proposed model was trained, validated, and tested on 234, 60, and 61 images, respectively. Across the entire dataset, the proposed method achieved the Dice Similarity Coefficient (DSC), 95th percentile Hausdorff Distance, sensitivity, precision, and number of breakage of 0.901±0.020, 3.275±5.885 mm, 0.912±0.26, 0.891±0.035, and 5.197±4.545. The sub-dataset-specific performance demonstrated consistent segmentation accuracy, with DSC values of 0.902±0.026, 0.893±0.032, 0.895±0.012, 0.911±0.008, 0.895±0.007, and 0.912±0.018, indicating minimal variability across the datasets. Furthermore, the proposed model outperformed both 3D Unet and SwinUnetr, delivering higher similarity to the ground truth and preserving the connectivity of the vascular structures more effectively. The average computational time of proposed model was 4.6 seconds.
Conclusion: The experimental results demonstrate that our model provides a fast and accurate solution for cerebrovascular segmentation. The proposed approach shows strong potential for further enhancement through the integration of robust and effective refinement techniques, paving the way for even greater reliability and precision in future applications.