CP-UNet: Contour-based Probabilistic Model for Medical Ultrasound Images Segmentation

Ruiguo Yu, Yiyang Zhang, Yuan Tian, Zhi-Qiang Liu, Xuewei Li, Jie Gao

TL;DR

CP-UNet is a novel segmentation framework designed to tackle the inherent blurriness and noise in medical ultrasound images. By modeling lesion contours as probabilistic distributions using Gaussian Mixture Models (GMMs) and employing a unique "channel-shift" downsampling technique, it significantly outperforms standard models in delineating tricky breast and thyroid nodules.

What Problem Does This Paper Address?

Ultrasound imaging is a cornerstone of modern diagnostics because it’s non-invasive and inexpensive. However, it’s notoriously "noisy." Ultrasound waves scatter and attenuate, leading to speckle noise, artifacts, and—most critically—contour blurring.

For a computer vision model, a blurry edge is a nightmare. Standard U-Net architectures often struggle because their convolutional kernels have a limited "receptive field" (they only look at small local patches), making it hard to distinguish a faint lesion boundary from the surrounding tissue. While some researchers have tried using Transformers to capture global context, these models can be computationally heavy and may overlook the sparse, fine-grained details of a contour.

Methods

The authors introduce CP-UNet, which incorporates three primary innovations:

  1. MgCSD (Multi-group Channel Shift Downsampling): Instead of traditional pooling or strided convolutions, this module divides channels into groups and "shifts" them spatially (upward and rightward). This creates global-local links early in the encoding process, allowing the network to see spatial relationships across the image without the massive overhead of a full self-attention mechanism.
  2. CPM (Contour Probabilistic Modeling): This is the "brain" of the approach. The authors treat the contour not as a fixed line, but as a probability distribution. They use a Gaussian Mixture Model (GMM) to fit the contour features. By using the "reparameterization trick" (commonly seen in Variational Autoencoders), the model can sample from these distributions to "predict" where a blurred contour likely exists based on learned global characteristics.
  3. GF (Gating-based Feature Filtering): This module acts as a smart fuser. It takes the up-sampled features, the skip-connection features, and the probabilistic contour features, then uses a gating mechanism to suppress noise and highlight the most relevant structural information.

Datasets

The model was evaluated on three distinct ultrasound datasets:

  • BUSI (Breast UltraSound Images): A public dataset containing 665 images with nodules.
  • DDTI (Dataset of Digital Thyroid Images): A public dataset with 872 images.
  • TUI: A large, private thyroid ultrasound dataset containing 15,233 images used to validate the results at scale.

Key Insights and Ideas

  • Embracing Uncertainty: Most segmentation models try to force a "hard" binary decision (is this pixel a lesion or not?). CP-UNet acknowledges that boundaries in ultrasound are inherently uncertain and uses GMMs to model that uncertainty mathematically.
  • Global Context at Shallow Levels: By using the MgCSD module, the network gains a "global" perspective even in its early layers. This is crucial because a blurred local edge can often only be resolved by looking at the broader shape of the lesion.
  • Breaking the Single-Loss Constraint: Traditional models rely on a single loss function (like Dice or Cross-Entropy). CP-UNet adds a KL Dispersion loss to align the predicted contour distribution with the actual ground truth distribution, providing a much stronger "guide" for the network during training.

Key Results

The quantitative results show that CP-UNet consistently beats both pure convolutional models (like UNet++) and Transformer-based models (like TransUNet):

  • Accuracy Boost: On the BUSI dataset, CP-UNet achieved an IoU of 0.6445 and a Dice score of 0.7827, outperforming the standard UNet by roughly 10-14%.
  • Superior Delineation: Qualitatively, the model excels at capturing "zigzag" or irregular contour variations. While even pre-trained TransUNet models tended to "smooth out" these intricate edges, CP-UNet followed the physician-labeled ground truth much more closely (see Fig. 4 in the paper).
  • Ablation Success: The authors proved that the MgCSD module alone improves the baseline by about 10%, but the best results only occur when all three modules (MgCSD, CPM, and GF) work in tandem.

Limitations and Future Work

While CP-UNet performs exceptionally well, the authors note that the CPM module is highly dependent on the quality of features provided by the downsampling stage. If the initial feature extraction is poor, the probabilistic modeling has less to work with.

For future work, the team aims to explore contour-based attention mechanisms to further improve diagnostic accuracy, potentially moving beyond just segmentation to help classify the "malignancy" of the detected nodules.

Summary updated

Abstract

Deep learning-based segmentation methods are widely utilized for detecting lesions in ultrasound images. Throughout the imaging procedure, the attenuation and scattering of ultrasound waves cause contour blurring and the formation of artifacts, limiting the clarity of the acquired ultrasound images. To overcome this challenge, we propose a contour-based probabilistic segmentation model CP-UNet, which guides the segmentation network to enhance its focus on contour during decoding. We design a novel down-sampling module to enable the contour probability distribution modeling and encoding stages to acquire global-local features. Furthermore, the Gaussian Mixture Model utilizes optimized features to model the contour distribution, capturing the uncertainty of lesion boundaries. Extensive experiments with several state-of-the-art deep learning segmentation methods on three ultrasound image datasets show that our method performs better on breast and thyroid lesions segmentation.
BibTeX
@misc{yu2024cp,
  author = {Ruiguo Yu and Yiyang Zhang and Yuan Tian and Zhi-Qiang Liu and Xuewei Li and Jie Gao},
  title = {{CP-UNet: Contour-based Probabilistic Model for Medical Ultrasound Images Segmentation}},
  year = {2024},
  eprint = {2411.14250},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2411.14250},
}
Figure