[Paper Review] IVUS-Net: An Intravascular Ultrasound Segmentation Network

Yang Ji, Tong Lin+2 more|arXiv (Cornell University)|Jun 10, 2018

Coronary Interventions and Diagnostics22 references7 citations

TL;DR

IVUS-Net is a fully convolutional neural network (FCN) with a post-processing contour extraction step that automatically segments both the lumen and media-adventitia boundaries in 20 MHz IVUS B-mode images. It achieves state-of-the-art performance, improving Hausdorff distance by 8% for the lumen and 20% for the media, while segmenting each frame in just 0.15 seconds on a modern GPU.

ABSTRACT

IntraVascular UltraSound (IVUS) is one of the most effective imaging modalities that provides assistance to experts in order to diagnose and treat cardiovascular diseases. We address a central problem in IVUS image analysis with Fully Convolutional Network (FCN): automatically delineate the lumen and media-adventitia borders in IVUS images, which is crucial to shorten the diagnosis process or benefits a faster and more accurate 3D reconstruction of the artery. Particularly, we propose an FCN architecture, called IVUS-Net, followed by a post-processing contour extraction step, in order to automatically segments the interior (lumen) and exterior (media-adventitia) regions of the human arteries. We evaluated our IVUS-Net on the test set of a standard publicly available dataset containing 326 IVUS B-mode images with two measurements, namely Jaccard Measure (JM) and Hausdorff Distances (HD). The evaluation result shows that IVUS-Net outperforms the state-of-the-art lumen and media segmentation methods by 4% to 20% in terms of HD distance. IVUS-Net performs well on images in the test set that contain a significant amount of major artifacts such as bifurcations, shadows, and side branches that are not common in the training set. Furthermore, using a modern GPU, IVUS-Net segments each IVUS frame only in 0.15 seconds. The proposed work, to the best of our knowledge, is the first deep learning based method for segmentation of both the lumen and the media vessel walls in 20 MHz IVUS B-mode images that achieves the best results without any manual intervention. Code is available at https://github.com/Kulbear/ivus-segmentation-icsm2018

Motivation & Objective

• To develop an end-to-end deep learning pipeline for automatic segmentation of both lumen and media-adventitia borders in IVUS images.
• To address the challenge of segmenting vessel walls in the presence of common artifacts such as bifurcations, shadows, and side branches.
• To achieve high-precision segmentation without relying on pre-trained weights or manual intervention.
• To enable real-time application by ensuring fast inference speed (0.15 seconds per frame).
• To establish a deep learning baseline for IVUS segmentation that surpasses traditional computer vision methods.

Proposed method

• Proposes a custom FCN architecture, IVUS-Net, trained from scratch without pre-trained ImageNet weights.
• Employs skip connections between encoder and decoder blocks to preserve spatial details and improve boundary localization.
• Introduces a refining branch to enhance feature representation and segmentation accuracy.
• Applies data augmentation techniques (e.g., rotation, scaling, flipping) to improve robustness and generalization.
• Uses a post-processing contour extraction step to refine the predicted pixel-wise masks into precise vessel wall contours.
• Trains and evaluates on a public IVUS benchmark dataset containing 326 end-diastolic frames with diverse artifacts.

Experimental results

Research questions

• RQ1 Can a fully convolutional network trained from scratch achieve superior segmentation performance for both lumen and media in IVUS images compared to state-of-the-art traditional methods?
• RQ2 How effective is the proposed IVUS-Net architecture in handling challenging IVUS artifacts such as bifurcations, shadows, and side branches not well-represented in the training set?
• RQ3 To what extent does data augmentation improve generalization and robustness of the segmentation model?
• RQ4 Does the inclusion of a refining branch significantly improve segmentation accuracy, particularly in terms of Hausdorff distance?
• RQ5 Can the proposed pipeline achieve real-time inference speed suitable for clinical deployment?

Key findings

• IVUS-Net outperforms the current state-of-the-art method (Faraji et al. [8]) by 4% in Jaccard Measure and 20% in Hausdorff Distance for media segmentation.
• The model improves Hausdorff Distance by 8% for lumen and 20% for media, indicating superior boundary localization accuracy.
• IVUS-Net achieves a Jaccard Measure of 0.82 (lumen) and 0.78 (media) on the test set, with Hausdorff Distance of 0.50 mm (lumen) and 0.82 mm (media).
• The model generalizes well to unseen artifacts: it performs strongly on test frames with bifurcations (44 frames), side vessels (93 frames), and shadows (96 frames).
• Data augmentation and the refining branch both contribute significantly to performance gains, as validated through ablation studies.
• The pipeline segments each IVUS frame in only 0.15 seconds on a modern GPU, enabling real-time clinical use.