DeepVesselNet: Vessel Segmentation, Centerline Prediction, and Bifurcation Detection in 3-D Angiographic Volumes (1803.09340v3)

Abstract: We present DeepVesselNet, an architecture tailored to the challenges faced when extracting vessel networks or trees and corresponding features in 3-D angiographic volumes using deep learning. We discuss the problems of low execution speed and high memory requirements associated with full 3-D convolutional networks, high-class imbalance arising from the low percentage of vessel voxels, and unavailability of accurately annotated training data - and offer solutions as the building blocks of DeepVesselNet. First, we formulate 2-D orthogonal cross-hair filters which make use of 3-D context information at a reduced computational burden. Second, we introduce a class balancing cross-entropy loss function with false positive rate correction to handle the high-class imbalance and high false positive rate problems associated with existing loss functions. Finally, we generate synthetic dataset using a computational angiogenesis model capable of generating vascular trees under physiological constraints on local network structure and topology and use these data for transfer learning. DeepVesselNet is optimized for segmenting and analyzing vessels, and we test the performance on a range of angiographic volumes including clinical MRA data of the human brain, as well as X-ray tomographic microscopy scans of the rat brain. Our experiments show that, by replacing 3-D filters with cross-hair filters in our network, we achieve over 23% improvement in speed, lower memory footprint, lower network complexity which prevents overfitting and comparable accuracy (with a Cox-Wilcoxon paired sample significance test p-value of 0.07 when compared to full 3-D filters). Our class balancing metric is crucial for training the network and transfer learning with synthetic data is an efficient, robust, and very generalizable approach leading to a network that excels in a variety of angiography segmentation tasks.

Overview of Deep VesselNet: Vessel Segmentation, Centerline Prediction, and Bifurcation Detection in 3-D Angiographic Volumes

The paper introduces Deep VesselNet, an advanced architecture aimed at tackling the complexities involved in segmenting vessel networks, predicting centerlines, and detecting bifurcations in 3-D angiographic volumes. This paper is crucial for enhancing the accuracy and efficiency of medical imaging tasks, specifically those requiring thorough analysis of vascular structures in modalities such as MRA and X-ray tomography. The authors address several notable challenges associated with three-dimensional convolutional networks: computational efficiency, class imbalance, and the scarcity of annotated training data.

Methodological Innovations

Deep VesselNet incorporates several notable innovations to optimize 3-D angiographic analysis:

1. Cross-hair Filters: The architecture replaces traditional 3-D convolutional filters with newly designed 2-D orthogonal cross-hair filters. This modification significantly reduces computational burden—improving execution speed by over 23%—without sacrificing accuracy. The alteration also reduces the risk of overfitting due to the lower complexity of the network.
2. Balance via Loss Function: Given the extreme class imbalance typical in such datasets, where vessel voxels are less than 3% of the total, the authors employ a class-balancing cross-entropy loss function coupled with a false positive rate correction strategy. This technique ensures numerical stability and curtails high false positive rates, resulting in more balanced precision-recall dynamics during training.
3. Synthetic Data and Transfer Learning: Recognizing the arduous nature of manually annotating 3-D vascular structures, the paper leverages synthetic datasets generated through a computational angiogenesis model for pretraining purposes. This model, validated through physiological constraints, generates vascular structures under realistic conditions, facilitating robust transfer learning that yields generalized network performance across different types of angiographic volumes.

Performance and Results

Empirical evaluations show Deep VesselNet's efficacy in various datasets, including clinical MRA and X-ray tomography. The results demonstrate that the network architecture, especially the Deep VesselNet-FCN variant sans subsampling layers, achieves superior Dice scores in vessel segmentation tasks. This variant excels due to its ability to retain fine-grained voxel details, which are critical for tasks involving curvilinear structures like vessels. Even though Deep VesselNet with cross-hair filters slightly underperforms compared to full 3-D convolutional networks in certain cases, the computational efficiency and memory savings present a pragmatic trade-off in large-scale clinical deployments.

Implications and Future Directions

The contributions of this paper have both theoretical and practical implications. On a theoretical level, the introduction of cross-hair filters enriches the workspace of convolutional operations, providing a viable alternative to full 3-D convolutions especially when constrained by memory and computational resources. Practically, the framework's application in diverse medical imaging modalities underscores its potential for enhancing clinical tools aimed at vascular diagnostics.

Looking ahead, the authors suggest extending Deep VesselNet toward simultaneous multiclass segmentation tasks that include vessels, their centerlines, and bifurcations within a unified framework. Such an advancement could streamline workflows in clinical angiographic analysis and foster further developments in deep learning applications in medical imaging.

In conclusion, Deep VesselNet represents a significant step towards automated, efficient analysis of complex vascular structures in 3-D imagery, paving the way for improved diagnostic accuracy in clinical settings. This work lays a robust foundation for subsequent explorations into specialized neural architectures adapted for nuanced medical imaging tasks.