TL-MSE²-Net: Transfer learning based nested model for cerebrovascular segmentation with aneurysms.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Neuroscience

Background:

• Cerebrovascular segmentation is crucial for diagnosing brain diseases.
• Current U-Net models struggle with tumor-affected vessels due to limited data and insufficient feature extraction.

Purpose of the Study:

• To develop an improved transfer learning model (TL-MSE²-Net) for enhanced cerebrovascular segmentation, particularly in cases with tumors.
• To address challenges of limited clinical datasets and inadequate feature extraction in existing models.

Main Methods:

• A transfer learning strategy using a pre-trained nested model (TL-MSE²-Net) based on 3D U-Net.
• Incorporation of ResMul, DeRes blocks for multiscale feature extraction and REAM block for edge voxel weighting.
• Leveraging publicly available datasets for pre-training to overcome data scarcity.

Main Results:

• The MSE²-Net framework achieved high Dice scores (70.81% and 89.08%) on public datasets, surpassing state-of-the-art methods.
• TL-MSE²-Net demonstrated superior performance on a private clinical dataset, with significant improvements in Dice score, sensitivity, Jaccard index, and precision.
• Ablation studies confirmed the effectiveness of individual model blocks.

Conclusions:

• The proposed TL-MSE²-Net effectively enhances cerebrovascular segmentation by improving feature extraction and leveraging transfer learning.
• This approach offers a promising solution for segmenting complex cerebrovascular structures, especially in the presence of tumors and limited data.