Clinical Report: Capsule Networks for Segmenting Small IVUS Image Datasets

Overview

This study evaluates the use of capsule networks for automatic segmentation of lumen and vessel wall in intravascular ultrasound (IVUS) images, focusing on small datasets. Capsule networks demonstrate potential advantages over conventional convolutional neural networks (CNNs) in handling limited training data due to their equivariant properties and part-whole hierarchical structure.

Background

Intravascular ultrasound (IVUS) imaging is essential for assessing vessel morphology and planning percutaneous coronary interventions by measuring parameters such as lumen diameter and vessel wall thickness. Manual delineation of these structures is time-consuming and operator-dependent, motivating the development of automatic segmentation methods. Traditional approaches include active contour models and machine learning techniques, while recent advances focus on CNNs, which require large annotated datasets. Capsule networks, a novel architecture encoding hierarchical relationships and pose information, may offer improved performance on small medical imaging datasets like IVUS.

Data Highlights

The study utilized two publicly available IVUS datasets: one with 435 annotated frames (384 × 384 pixels) from 10 patients acquired at 20 MHz, and another with 77 images (512 × 512 pixels) from 22 patients acquired at 40 MHz. Annotations included lumen border and external elastic membrane delineations transformed into pixel masks for lumen, vessel wall, and background. Capsule networks were optimized and compared to state-of-the-art CNNs for segmentation performance, particularly focusing on robustness with limited training data.

Key Findings

• Capsule networks encode entities and their properties (pose, texture, deformation) via groups of neurons called capsules, enabling part-whole hierarchical relationships through iterative routing.
• Capsule networks exhibit equivariance not only to translations but also to complex transformations such as rotations and reflections, potentially improving generalizability on small datasets.
• Compared to CNNs, capsule networks can better handle small IVUS datasets due to their ability to learn pose-aware representations and parse tree-like structures.
• The optimized capsule network architecture tailored for ultrasound image characteristics (e.g., speckle noise) achieved competitive segmentation of lumen and vessel wall.
• Segmentation performance depends on dataset quality and size; capsule networks reduce dependency on large annotated datasets by leveraging intrinsic image structure.

Clinical Implications

Automatic segmentation of IVUS images using capsule networks can streamline vessel morphology assessment, reducing manual workload and inter-operator variability. Their robustness on small datasets makes them suitable for clinical settings where expert annotations are limited. This approach may enhance treatment planning efficiency for percutaneous coronary interventions by providing reliable vessel parameter measurements.

Conclusion

Capsule networks represent a promising alternative to CNNs for IVUS image segmentation, particularly when training data are scarce. Their ability to model hierarchical relationships and pose information improves segmentation robustness and may facilitate clinical adoption of automated IVUS analysis.

References

1. LaLonde et al. 2018 -- Capsule Networks for Small Dataset Segmentation
2. Sabour et al. 2017 -- Dynamic Routing Between Capsules
3. Zhou et al. 2019 -- U-Net: Convolutional Networks for Biomedical Image Segmentation
4. Molinari et al. 2020 -- IVUS Segmentation Using Machine Learning