Multi-channel multi-scale fully convolutional network for 3D perivascular spaces segmentation in 7T MR images

Area of Science:

• Medical Imaging
• Neuroscience
• Artificial Intelligence

Background:

• Accurate segmentation of perivascular spaces (PVSs) is crucial for quantitative analysis of PVS morphology.
• Manual delineation of PVSs is challenging due to their thin, low-contrast tubular structures and large numbers.
• Existing automated methods often rely on hand-crafted features and complex preprocessing, limiting their generalizability.

Purpose of the Study:

• To develop a novel fully convolutional neural network (FCN) for efficient and accurate segmentation of PVSs.
• To eliminate the need for hand-crafted features and predefined regions-of-interest (ROIs).
• To improve the quantitative study of PVS morphology.

Main Methods:

• A multi-channel, multi-scale FCN was proposed, utilizing both original and enhanced T2-weighted 7T MR images.
• A non-local Haar-transform-based line singularity representation enhanced tubular structures.
• Recursive input of PVS probability maps refined segmentation by providing auxiliary contextual information.

Main Results:

• The proposed FCN demonstrated superior performance in PVS segmentation compared to state-of-the-art methods.
• The method effectively utilizes multi-channel inputs for comprehensive image information.
• Automatic learning of multi-scale features improved characterization of PVS spatial associations.

Conclusions:

• The novel FCN offers an efficient and accurate solution for PVS segmentation in 7T brain MR images.
• The approach overcomes limitations of traditional methods by avoiding manual feature engineering and ROIs.
• This method advances quantitative PVS morphology studies.