Shape-Sensitive Loss for Catheter and Guidewire Segmentation

TL;DR

This work addresses the challenge of precisely segmenting catheters and guidewires in X-ray imagery by introducing a shape-sensitive loss that operates on Signed Distance Maps (SDMs) and leverages a Vision Transformer (ViT) to capture global boundary geometry. Predictions and ground-truth labels are converted into SDMs, from which high-level ViT features are extracted and compared via cosine similarity, formalized as $\mathcal{L}_{SS}=1-\text{CosSim}(\text{ViT}(\text{SDM}_{Output}),\text{ViT}(\text{SDM}_{Label}))$ with $\text{CosSim}(\alpha,\beta)=\dfrac{\alpha\cdot\beta}{|\alpha||\beta|}$, and combined with Dice loss as $\mathcal{L}_{total}=\gamma\mathcal{L}_{Dice}+\delta\mathcal{L}_{SS}$. The approach is validated on a large dataset of real-animal and phantom X-ray images across multiple backbones (e.g., TransU-Net, SwinU-Net), yielding consistent improvements in Dice, IoU, and boundary delineation and establishing state-of-the-art performance. Results show that cosine-based feature distances in SDM space outperform traditional distance measures, and the Dice+shape-sensitive loss framework enhances segmentation without increasing model complexity. The findings suggest a practical boundary-aware, scale- and translation-invariant paradigm for critical instrument delineation in interventional imaging, with potential applicability to other boundary-sensitive medical segmentation tasks.

Abstract

We introduce a shape-sensitive loss function for catheter and guidewire segmentation and utilize it in a vision transformer network to establish a new state-of-the-art result on a large-scale X-ray images dataset. We transform network-derived predictions and their corresponding ground truths into signed distance maps, thereby enabling any networks to concentrate on the essential boundaries rather than merely the overall contours. These SDMs are subjected to the vision transformer, efficiently producing high-dimensional feature vectors encapsulating critical image attributes. By computing the cosine similarity between these feature vectors, we gain a nuanced understanding of image similarity that goes beyond the limitations of traditional overlap-based measures. The advantages of our approach are manifold, ranging from scale and translation invariance to superior detection of subtle differences, thus ensuring precise localization and delineation of the medical instruments within the images. Comprehensive quantitative and qualitative analyses substantiate the significant enhancement in performance over existing baselines, demonstrating the promise held by our new shape-sensitive loss function for improving catheter and guidewire segmentation.

Figures (4)

• Figure 1: Catheter and guidewrite segmentation in X-ray images. First row: The input X-ray images. Second row: The segmentation results. Red color denotes the catheter, green color denotes the guidewire.
• Figure 2: Illustration of the process to create the Signed Distance Maps. Top Row: Original groundtruth images. Bottom Row: Signed Distance Maps, calculated based on the contours, overlaid by its contour images
• Figure 3: An overview of our framework.
• Figure 4: Effects of Blending Parameters: $\gamma$ for Dice Loss and $\delta$ for Our Proposed Shape-Sensitive Loss on Various Network Architectures.