KaLDeX: Kalman Filter based Linear Deformable Cross Attention for Retina Vessel Segmentation

Zhihao Zhao, Shahrooz Faghihroohi, Yinzheng Zhao, Junjie Yang, Shipeng Zhong, Kai Huang, Nassir Navab, Boyang Li, M. Ali NasseriarXiv 2024 · DOI · arXiv

Introduction: The Window to Systemic Health

In ophthalmology, the retinal vasculature—the intricate network of blood vessels at the back of the eye—serves as a unique, non-invasive window into a patient's systemic health. Pathological changes in these vessels, such as narrowing, leakage, or abnormal branching, are early indicators of chronic conditions including diabetes, hypertension, and various ocular diseases. Consequently, precise segmentation of these vessels from medical images is a cornerstone of automated diagnostic systems.

While deep learning models, particularly the U-Net architecture and its variants, have set high standards for medical image segmentation, they face a persistent challenge known as the "small vessel problem." Standard convolutional neural networks (CNNs) rely on successive downsampling layers to extract high-level semantic features. However, this process often obliterates the fine, high-resolution details of minuscule vessels, leading to fragmented or entirely missing segments in the resulting masks. Furthermore, the rigid, square kernels of traditional CNNs are fundamentally ill-suited to the curvilinear, snake-like geometry of the human vascular system.

Methodology: Tracking Vessels as Trajectories

To overcome these limitations, the authors propose KaLDeX (Kalman Filter based Linear Deformable X-Attention). This architecture moves away from rigid pixel-wise classification and instead treats the path of a vessel as a trajectory to be tracked and refined.

The KaLDeX architecture (a) utilizes a UNet++ backbone integrated with Linear Deformable Cross Attention (LDCA) modules (b) to capture both local vessel details and global context.

Linear Deformable (LD) Module

The core of the KaLDeX approach is the Linear Deformable Module. Inspired by the concept of "snake convolutions," this module allows the convolutional kernel to dynamically adjust its sampling locations. Instead of a fixed grid, the kernel can stretch and bend to align with the underlying vessel structure, effectively "following the curves" of the anatomy.

Kalman Filter Integration

The "secret sauce" of the paper is the integration of a Kalman Filter—an algorithm traditionally used in GPS and radar systems for state estimation—to optimize the deformable kernel's offsets. In standard deformable convolutions, the predicted offsets can be noisy, causing the kernel to "drift" off the vessel path. By treating the sequence of offsets as a trajectory, the Kalman Filter smooths these predictions, filtering out noise and ensuring the kernel remains precisely locked onto the vascular centerline.

Linear Deformable Cross-Attention (LDCA)

To ensure that local structural details are not lost in the broader context of the image, the authors introduce the LDCA module. This module acts as a bridge, using cross-attention to fuse high-resolution local features from the LD module with global contextual features from the UNet++ backbone. This dual-stream approach ensures that even the thinnest capillaries are recognized as part of the larger vascular tree.

Topological Loss and Connectivity

Beyond pixel-level accuracy, the model prioritizes structural integrity through a topological loss function based on persistent homology. This mathematical framework penalizes the model for creating "breaks" or "holes" in what should be continuous vessel branches. By optimizing for topology, KaLDeX ensures that the resulting segmentations are physically plausible and clinically useful.

Datasets and Evaluation

The robustness of KaLDeX was rigorously tested across five major clinical benchmarks:

  • Fundus Images: The classic DRIVE, CHASE_DB1, and STARE datasets.
  • OCTA-500: High-resolution Optical Coherence Tomography Angiography images, which provide a much more detailed view of the microvasculature than standard fundus photography.
  • FIVES: A large-scale dataset of 800 high-resolution images categorized by disease state (e.g., AMD, Glaucoma, Diabetic Retinopathy).

Key Results and Discussion

KaLDeX established new state-of-the-art (SOTA) performance across all metrics, with a particular emphasis on connectivity.

  • Pixel Accuracy: The model achieved an accuracy of 97.25% on DRIVE and 97.85% on STARE, consistently outperforming recent heavyweights like TransUNet and FR-UNet.
  • Structural Integrity: On the clDice (Centerline Dice) metric, which specifically measures how well the "skeleton" of the vessels is preserved, KaLDeX outperformed baselines significantly, reaching 94.02% on the complex OCTA-500 dataset.
  • Data Efficiency: The authors demonstrated that KaLDeX is remarkably robust in "few-shot" scenarios. Even when trained on as few as five images, the model maintains a coherent understanding of vascular structures, a vital trait in medical domains where expert-annotated data is a scarce resource.

Conclusion

KaLDeX represents a significant advancement in medical image segmentation by successfully merging classical signal processing (Kalman filtering) with modern deep learning (deformable convolutions and cross-attention). By prioritizing the structural continuity of the retinal vasculature, it provides a more reliable tool for the early detection and monitoring of systemic diseases.

Summary updated

BibTeX
@misc{zhao2024kaldex,
  author = {Zhihao Zhao and Shahrooz Faghihroohi and Yinzheng Zhao and Junjie Yang and Shipeng Zhong and Kai Huang and Nassir Navab and Boyang Li and M. Ali Nasseri},
  title = {{KaLDeX: Kalman Filter based Linear Deformable Cross Attention for Retina Vessel Segmentation}},
  year = {2024},
  eprint = {2410.21160},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2410.21160},
}
Figure