Benefit from public unlabeled data: A Frangi filtering-based pretraining network for 3D cerebrovascular segmentation

TL;DR

This paper introduces a Frangi filtering-based pretraining method leveraging a large unlabeled TOF-MRA dataset to improve 3D cerebrovascular segmentation, achieving superior results over existing semi- and self-supervised approaches.

Contribution

Proposes a novel pretraining strategy using Frangi filtering on large unlabeled data to enhance cerebrovascular segmentation performance.

Findings

Achieved ~3% improvement over state-of-the-art methods.

Demonstrated effectiveness and generalizability of the pretraining approach.

Provided large-scale unlabeled dataset and open-source code.

Abstract

The precise cerebrovascular segmentation in time-of-flight magnetic resonance angiography (TOF-MRA) data is crucial for clinically computer-aided diagnosis. However, the sparse distribution of cerebrovascular structures in TOF-MRA results in an exceedingly high cost for manual data labeling. The use of unlabeled TOF-MRA data holds the potential to enhance model performance significantly. In this study, we construct the largest preprocessed unlabeled TOF-MRA datasets (1510 subjects) to date. We also provide three additional labeled datasets totaling 113 subjects. Furthermore, we propose a simple yet effective pertraining strategy based on Frangi filtering, known for enhancing vessel-like structures, to fully leverage the unlabeled data for 3D cerebrovascular segmentation. Specifically, we develop a Frangi filtering-based preprocessing workflow to handle the large-scale unlabeled dataset, and a multi-task pretraining strategy is proposed to effectively utilize the preprocessed data. By employing this approach, we maximize the knowledge gained from the unlabeled data. The pretrained model is evaluated on four cerebrovascular segmentation datasets. The results have demonstrated the superior performance of our model, with an improvement of approximately 3\% compared to state-of-the-art semi- and self-supervised methods. Furthermore, the ablation studies also demonstrate the generalizability and effectiveness of the pretraining method regarding the backbone structures. The code and data have been open source at: \url{https://github.com/shigen-StoneRoot/FFPN}.