Ultrasound thyroid segmentation based on segmented frequency domain and local attention

Yishan Hu, Pinle Qin, Jianchao Zeng, Rui Chai, Lifang WangJournal of Image and Graphics 2020 · DOI

TL;DR

Researchers have developed a novel deep learning architecture that combines traditional frequency-domain signal processing with modern attention mechanisms to accurately segment thyroid glands in ultrasound images. By explicitly handling the "speckle noise" inherent in ultrasound and focusing on multi-scale features, the model achieves a Dice coefficient of 0.812, significantly outperforming standard benchmarks like U-Net.

What Problem Does This Paper Address?

Thyroid segmentation is a critical step in diagnosing endocrine diseases and planning surgeries, as volume and shape are key clinical indicators. However, ultrasound (US) imaging—the primary tool for thyroid examination—is notoriously difficult for computer vision models.

The authors identify three main hurdles:

  1. Speckle Noise: The random intensity patterns (speckle) in US images obscure anatomical boundaries.
  2. Low Contrast: Thyroid tissue often looks very similar to surrounding muscles and glands.
  3. Morphological Variance: The size and shape of the thyroid vary wildly between patients and different imaging hardware.

Existing deep learning models often struggle to generalize across different data sources or fail to capture the fine-grained edges needed for precise clinical measurement.

Methods

The authors propose a hybrid network based on an encoder-decoder structure (using a ResNet34 backbone) enhanced by three key modules:

  • Frequency Domain Feature Augmentation (FFA): Instead of feeding raw pixels alone, the model uses high-pass and low-pass filters to extract edge details and structural outlines. This "traditional" image processing step helps the network ignore random speckle noise and focus on the underlying anatomy.
  • Local Attention Mechanism (LA): This module consists of two parts:
  • Channel Attention: Identifies which feature maps (e.g., edges vs. textures) are most important.
  • Spatial Attention: Identifies where in the image the target is, helping the model ignore complex non-target tissues.
  • Cascaded Multi-scale Dilated Convolution Pyramid (CMDP): By using dilated convolutions with varying rates, the model expands its "receptive field" without losing resolution. This allows it to recognize thyroid structures regardless of whether they appear large or small in the frame.
  • Hybrid Loss Function: The researchers combined Binary Cross Entropy (BCE) for pixel-level accuracy with Structural Similarity (SSIM) loss to ensure the predicted contours match the actual anatomical boundaries.

Datasets

The study utilized a publicly available dataset consisting of 3,424 thyroid ultrasound images collected from 16 healthy individuals using 11–16 MHz probes. The ground truth labels were hand-drawn by experts. To ensure robustness, the authors employed 10-fold cross-validation and standard data augmentation techniques like random horizontal flipping.

Key Insights and Ideas

  • Traditional Meets Modern: The biggest "aha" moment is that deep learning doesn't have to do everything from scratch. By using frequency-domain filtering (FFA) as a preprocessing step within the network, the authors effectively "cleaned" the data for the convolutional layers, making the segmentation task much easier.
  • Attention as a Noise Filter: In ultrasound, "noise" isn't just random pixels; it's complex anatomical structures that look like the target. The Local Attention module acts as a cognitive filter, forcing the model to "look away" from the carotid artery or neck muscles and focus on the thyroid.
  • Contour Matters: Most segmentation models just try to get the pixels right. By adding SSIM to the loss function, the authors prioritized the shape and smoothness of the boundary, which is more clinically relevant than just pixel-wise accuracy.

Key Results

The proposed model was compared against several heavy hitters in medical imaging:

Method Accuracy Recall Precision Dice Coefficient
U-Net 0.962 0.793 0.911 0.751
CE-Net 0.979 0.841 0.932 0.798
Proposed Model 0.989 0.849 0.941 0.812

The model showed a marked improvement in the Dice coefficient (a measure of overlap between the prediction and the ground truth), indicating much more reliable boundary detection. Qualitative results showed the model was particularly good at maintaining smooth contours even when the ultrasound signal was weak or noisy.

Limitations and Future Work

Despite the strong performance, the authors noted that the model still struggles with extremely small thyroid structures. When the target is tiny, the complex surrounding tissues can still cause "false negatives" or missegmentation.

Future research directions include:

  • Integrating multi-scale object detection modules specifically designed for small objects.
  • Exploring boundary detection modules to further refine the edge precision.
  • Testing the model on larger, more diverse clinical datasets to ensure it handles pathological cases (like nodules or goiters) as well as it handles healthy tissue.

Summary updated

Abstract

目的超声检查是诊断甲状腺疾病的主要影像学方法之一,但由于超声图像中斑点强度具有随机性、组织器官复杂等问题,导致甲状腺在不同数据源间的形态、大小和纹理差异性较大,容易导致观察者视觉疲劳。针对甲状腺超声成像存在斑点强度随机性以及周边组织复杂性的问题,为了更准确地描述出器官与病理性病变的解剖边界,提出一种基于频域增强和局部注意力机制的甲状腺超声分割网络。方法针对原始数据采用高低通滤波器获取高低频段的图像信息,整合高频段细节特征与低频段边缘特征,增强图像前背景的对比度,降低图像间的差异性。根据卷积网络中网络深度所提取特征信息量的不同,采用局部注意力机制对高低维特征信息进行自适应激活,增强低维特征的细节信息,弱化对非目标区域的关注,增强高维特征的全局信息,弱化冗余信息对网络的干扰,增强前背景分类以及对非显著性目标检测的能力。采用金字塔级联空洞卷积获取不同感受野的特征信息,解决数据源间图像差异较大的问题。结果实验结果表明,本文方法在11~16 MHz时采集的16个手绘甲状腺超声公开数据集中,通过10折交叉验证显示准确率为0.989,召回率为0.849,精准率为0.940,Dice系数为0.812,效果优于当前其他医学图像分割网络。通过消融实验,证明本文的几个模块对超声图像分割确实具有一定的提升效果。结论本文所提分割网络,结合深度学习模型及传统图像处理模型的优点,能较好地处理超声图像随机斑点并且提升非显著性组织分割效果。
BibTeX
@article{hu2020ultrasound,
  author = {Yishan Hu and Pinle Qin and Jianchao Zeng and Rui Chai and Lifang Wang},
  title = {{Ultrasound thyroid segmentation based on segmented frequency domain and local attention}},
  journal = {Journal of Image and Graphics},
  year = {2020},
  doi = {10.11834/jig.200230},
}
Figure