[Paper Review] SCOPE: Structural Continuity Preservation for Medical Image Segmentation

Yousef Yeganeh, Azade Farshad+6 more|arXiv (Cornell University)|Apr 28, 2023

Retinal Imaging and AnalysisMedicine3 citations

TL;DR

This paper proposes SCOPE, a graph-based deep learning method that enforces structural continuity in medical image segmentation by modeling anatomical topology as a graph constraint. By treating image patches as nodes and leveraging graph convolutional networks with clDice loss, SCOPE significantly improves connectivity metrics—reducing topology errors by up to 90%—while maintaining state-of-the-art pixel-wise segmentation performance on retinal vessel datasets.

ABSTRACT

Although the preservation of shape continuity and physiological anatomy is a natural assumption in the segmentation of medical images, it is often neglected by deep learning methods that mostly aim for the statistical modeling of input data as pixels rather than interconnected structures. In biological structures, however, organs are not separate entities; for example, in reality, a severed vessel is an indication of an underlying problem, but traditional segmentation models are not designed to strictly enforce the continuity of anatomy, potentially leading to inaccurate medical diagnoses. To address this issue, we propose a graph-based approach that enforces the continuity and connectivity of anatomical topology in medical images. Our method encodes the continuity of shapes as a graph constraint, ensuring that the network's predictions maintain this continuity. We evaluate our method on two public benchmarks on retinal vessel segmentation, showing significant improvements in connectivity metrics compared to traditional methods while getting better or on-par performance on segmentation metrics.

Motivation & Objective

• To address the limitation of deep learning models in preserving anatomical continuity during medical image segmentation.
• To model the topological structure of biological structures like blood vessels as a graph constraint to enforce connectivity.
• To improve segmentation performance not only in pixel-wise accuracy but also in topological fidelity, especially for tubular structures.
• To develop a method that is agnostic to existing CNN architectures and integrates seamlessly with them.

Proposed method

• The input image is divided into non-overlapping n×n patches, which are treated as nodes in a graph to encode spatial topology.
• Graph convolutional networks (GCNs) are used to propagate features across neighboring patches, enforcing structural continuity through message passing.
• Node features are generated from patch embeddings using a backbone network (DRIU), and graph features are fused with spatial features via skip connections.
• A hybrid loss function combining cross-entropy and clDice is used to optimize both pixel-wise accuracy and connectivity.
• The graph structure is built using 1-level edge connections between adjacent patches, ensuring local neighborhood consistency.
• The method is architecture-agnostic and can be integrated with existing U-Net or attention-based networks like DRIU.

Experimental results

Research questions

• RQ1 Can a graph-based approach effectively preserve anatomical continuity in medical image segmentation without relying on architectural priors?
• RQ2 How does enforcing topological constraints via graph convolution improve connectivity metrics compared to standard deep learning models?
• RQ3 To what extent does the choice of patch size and loss function affect the preservation of vessel continuity in retinal images?
• RQ4 Can the proposed method outperform state-of-the-art models in both segmentation accuracy and topological fidelity on public retinal vessel datasets?

Key findings

• On the FIVES dataset, SCOPE reduced the $error_{\beta_0}$ metric by 90% compared to DRIU (from 73.8 to 12.1), indicating superior topological preservation.
• SCOPE achieved a clDice score of 0.94 on the AMD class, outperforming DRIU (0.87) and RU-UNet (0.87), demonstrating improved connectivity modeling.
• The ablation study showed that using clDice loss alone reduced $error_{\beta_0}$ to 12.1, while CE loss alone led to 52.6, highlighting its critical role in connectivity.
• Patch size analysis revealed that 2×2 patches degraded performance (e.g., $error_{\beta_0}$ = 12.83), suggesting optimal patch size is 1×1 for centerline preservation.
• SCOPE maintained competitive pixel-wise metrics (Dice: 0.85–0.91) across all classes, showing no trade-off with segmentation accuracy.
• Qualitative results confirmed that SCOPE preserved vessel continuity where baselines introduced spurious disconnections, especially in thin or tortuous vessels.