Cross-modality cerebrovascular segmentation based on pseudo-label generation via paired data

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Accurate segmentation of cerebrovascular structures is vital for diagnosing cranial vascular diseases using medical imaging modalities like CTA, MRA, and DSA.
• Deep Convolutional Neural Networks (CNNs) have advanced medical image segmentation, but require extensive, costly, and time-consuming data labeling for each imaging type.

Purpose of the Study:

• To develop a cross-modality cerebrovascular segmentation network that minimizes the need for manual data labeling across different imaging modalities.
• To leverage paired data from source and target domains for efficient training of segmentation networks.

Main Methods:

• A universal vessel segmentation network was trained using manually labeled source domain data.
• Initial labels for target domain training images were automatically generated.
• Paired images were fused to refine the initial labels, which then trained the target domain segmentation network.

Main Results:

• The proposed method demonstrated effectiveness in cross-modality cerebrovascular segmentation.
• Experiments on MRA-CTA and DSA-CTA datasets confirmed state-of-the-art performance.
• The approach successfully reduced the reliance on extensive manual annotation for each imaging modality.

Conclusions:

• The developed method offers an efficient solution for training cross-modality cerebrovascular segmentation networks.
• This technique has the potential to significantly streamline the diagnostic process for cranial vascular diseases.
• The approach achieves high performance, comparable to state-of-the-art methods, while reducing data acquisition and labeling costs.