A nested parallel multiscale convolution for cerebrovascular segmentation.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Cerebrovascular segmentation in MRI is crucial for diagnosing and treating cerebrovascular diseases.
• Existing convolutional neural network (CNN) and U-Net-like frameworks struggle with segmenting small or thin cerebrovascular structures.
• Limitations include inadequate attention to multiscale features, limited feature extraction depth, and suboptimal feature utilization in decoders.

Purpose of the Study:

• To propose a novel 3D U-Net-like framework, Usception, specifically designed for improved small cerebrovascular segmentation.
• To address limitations in existing methods by enhancing multiscale feature extraction and utilization.
• To improve the accuracy of cerebrovascular segmentation in MRI.

Main Methods:

• Developed Usception, a 3D U-Net-like framework inspired by U-Net++.
• Incorporated three key blocks: Reduction block, Gap block, and Deep block.
• Utilized grouped convolutions of different sizes and varied path depths between encoder and decoder to enhance multiscale feature processing.

Main Results:

• Evaluated Usception on three public and in-house clinical MRA datasets.
• Achieved average Dice scores of 69.29%, 87.40%, and 77.77% across the datasets.
• Demonstrated superior performance compared to existing state-of-the-art methods, with ablation studies validating block effectiveness.

Conclusions:

• The Usception framework effectively maximizes multiscale feature extraction and utilization.
• Combines Inception-ResNet principles with a dimension-expanded U-Net++ architecture.
• Achieves competitive and improved segmentation results for small cerebrovascular structures.