SPGNet: A Shape-prior Guided Network for Medical Image Segmentation

Zhengxuan Song, Xun Liu, Wenhao Zhang, Yongyi Gong, Tianyong Hao, Kun ZengIJCAI 2024

Abstract

Medical image segmentation often struggles with global anatomical consistency, leading to fragmented or biologically implausible results. We present SPGNet (Shape-prior Guided Network), a dual-path architecture that combines pixel-level deep learning with classical statistical shape modeling. By explicitly learning anatomical shape distributions and employing a Transformer-based refinement module, SPGNet achieves state-of-the-art performance on chest X-ray and breast ultrasound datasets, producing masks with high anatomical integrity.

Introduction

Deep learning models like U-Net have revolutionized medical image segmentation through high-precision local feature extraction. However, these models often lack a global understanding of anatomy, which can result in "leaking" boundaries or fragmented shapes in low-contrast or noisy images. While shape priors have been explored previously, they often lack the flexibility to handle significant inter-patient variability. SPGNet addresses this by integrating flexible statistical shape priors directly into the neural network's learning process, bridging the gap between "old school" Active Shape Models (ASM) and modern deep learning.

Methodology

The SPGNet architecture consists of two primary paths that collaborate to produce refined segmentation masks.

Overview of the SPGNet architecture, featuring the Segmentation path and the Shape-prior path connected via Dual-path Collaboration Modules (DCM).

Dual-Path Architecture

  1. Segmentation Path: A standard encoder-decoder structure focused on dense, pixel-wise classification.
  2. Shape-prior Path: This path utilizes offline multi-class shape modeling (via Procrustes analysis and PCA) to classify input images into shape categories and regress deformation parameters, effectively fitting a statistical "skeleton" to the anatomy.

Collaboration and Refinement

The Dual-path Collaboration Module (DCM) facilitates communication between the two paths using spatial and channel attention. This allows the segmentation path to benefit from structural guidance while providing the shape path with rich contextual features. Finally, a Long-range Refinement Module (LRM), based on a Transformer architecture, crops local features around predicted shape points and applies self-attention to refine their positions, ensuring global consistency.

Experiments and Results

The model was evaluated on the JSRT (chest X-ray) and BUS (breast ultrasound) datasets.

Quantitative Performance

On the JSRT dataset, SPGNet achieved a Dice score of 98.01% for lung segmentation, significantly improving boundary accuracy as measured by the 95% Hausdorff Distance (HD95). On the noisy BUS dataset, it reached a Dice score of 78.40%, outperforming previous state-of-the-art models like AAUNet (77.68%). This demonstrates the model's ability to maintain anatomical integrity even when the image signal is weak or noisy.

Qualitative Analysis

Qualitative results demonstrate that SPGNet produces smoother, more anatomically correct boundaries compared to models without shape-prior guidance. The attention maps in deeper layers clearly shift toward edges, which are the areas most susceptible to segmentation errors.

Qualitative comparison of segmentation results. The inclusion of shape priors (w/ s-p) leads to more anatomically plausible predictions with fewer artifacts compared to versions without (w/o s-p).

Discussion and Conclusion

SPGNet successfully bridges the gap between interpretable statistical shape models and the representational power of deep learning. By providing the network with a "mental model" of anatomy, it achieves segmentations that are both mathematically accurate and biologically sound. While the dual-path design increases computational complexity, the gains in anatomical integrity represent a significant step forward for reliable medical AI. Future work includes extending this approach to 3D volumetric data such as CT or MRI scans.

Summary updated

BibTeX
@inproceedings{song2024spgnet,
  author = {Zhengxuan Song and Xun Liu and Wenhao Zhang and Yongyi Gong and Tianyong Hao and Kun Zeng},
  title = {{SPGNet: A Shape-prior Guided Network for Medical Image Segmentation}},
  booktitle = {International Joint Conference on Artificial Intelligence},
  year = {2024},
}
Figure