Fully-automatic segmentation of kidneys in clinical ultrasound images using a boundary distance regression network

Shi Yin, Zhengqiang Zhang, Hongming Li, Qinmu Peng, Xinge You, Susan L. Furth, Gregory E. Tasian, Yong Fan

Abstract

It remains challenging to automatically segment kidneys in clinical ultrasound images due to the kidneys' varied shapes and image intensity distributions, although semi-automatic methods have achieved promising performance. In this study, we developed a novel boundary distance regression deep neural network to segment the kidneys, informed by the fact that the kidney boundaries are relatively consistent across images in terms of their appearance. Particularly, we first use deep neural networks pre-trained for classification of natural images to extract high-level image features from ultrasound images, then these feature maps are used as input to learn kidney boundary distance maps using a boundary distance regression network, and finally the predicted boundary distance maps are classified as kidney pixels or non-kidney pixels using a pixel classification network in an end-to-end learning fashion. Experimental results have demonstrated that our method could effectively improve the performance of automatic kidney segmentation, significantly better than deep learning based pixel classification networks.

Introduction

Accurate kidney segmentation in ultrasound (US) imaging is critical for diagnosing conditions like chronic kidney disease and congenital abnormalities of the kidney and urinary tract (CAKUT). However, ultrasound images are notoriously difficult to process automatically due to high variability in organ shape and size, speckle noise, and intensity overlap with surrounding tissues. While semi-automatic methods exist, they require manual initialization, which is slow and prone to human error. Standard deep learning models often struggle because they treat every pixel inside the kidney as identical, ignoring the structural importance of the organ's boundary. This paper introduces a framework that shifts the focus from simple pixel-wise classification to regressing boundary distance maps, leveraging the observation that kidney boundaries are more consistent in appearance than their interiors.

Methodology

The proposed framework utilizes a two-stage, end-to-end deep learning pipeline:

  1. Feature Extraction: To overcome the lack of massive medical datasets, a VGG-16 network pre-trained on ImageNet is used. It is adapted with atrous (dilated) convolutions to extract dense, high-level features while maintaining spatial resolution.
  2. Boundary Distance Regression Network (BRN): Instead of immediate binary mask prediction, this network learns to predict a normalized exponential distance map. In this map, pixels on the boundary have a value of 1, and values decrease exponentially as distance from the boundary increases. This forces the model to learn the consistent structural features of kidney edges.
  3. Pixel Classification Network: The predicted distance map is fed into a final classification layer that produces the binary mask.
Illustration of the boundary distance map concept, where distance from the kidney boundary is mapped to a normalized exponential value.

The model is trained using a dynamic loss weighting strategy that prioritizes boundary detection early in the training process before shifting focus toward final pixel classification and mask refinement.

Results and Discussion

The study utilized a clinical dataset from the Children’s Hospital of Philadelphia (CHOP) comprising 185 ultrasound images from 50 normal subjects and 50 patients with CAKUT. The proposed method demonstrated a marked improvement over established semantic segmentation models like FCN and DeepLab.

Method Dice Coefficient Mean Distance (pixels) Accuracy
FCN 0.78 ± 0.11 9.83 ± 4.82 0.96 ± 0.02
DeepLab 0.86 ± 0.11 5.85 ± 3.28 0.97 ± 0.02
Proposed 0.94 ± 0.03 2.72 ± 1.61 0.989 ± 0.006

Qualitatively, the model successfully segmented kidneys even in heterogeneous cases where the organ's appearance was blurred or distorted by disease. The system is highly efficient, taking only 0.18 seconds to segment a single image on a consumer-grade GPU.

Qualitative comparison of kidney segmentation results across different methods: Manual ground truth, FCN, DeepLab, and the proposed Boundary Distance Regression Network.

Conclusion

By regressing a continuous distance map rather than performing direct binary classification, the proposed network provides a more robust signal for optimization and better captures the structural consistency of kidney boundaries. Future work will focus on extending this logic to 3D ultrasound volumes and validating the model on larger, multi-center datasets.

Summary updated

Abstract

It remains challenging to automatically segment kidneys in clinical ultrasound images due to the kidneys' varied shapes and image intensity distributions, although semi-automatic methods have achieved promising performance. In this study, we developed a novel boundary distance regression deep neural network to segment the kidneys, informed by the fact that the kidney boundaries are relatively consistent across images in terms of their appearance. Particularly, we first use deep neural networks pre-trained for classification of natural images to extract high-level image features from ultrasound images, then these feature maps are used as input to learn kidney boundary distance maps using a boundary distance regression network, and finally the predicted boundary distance maps are classified as kidney pixels or non-kidney pixels using a pixel classification network in an end-to-end learning fashion. Experimental results have demonstrated that our method could effectively improve the performance of automatic kidney segmentation, significantly better than deep learning based pixel classification networks.
BibTeX
@inproceedings{yin2019fully,
  author = {Shi Yin and Zhengqiang Zhang and Hongming Li and Qinmu Peng and Xinge You and Susan L. Furth and Gregory E. Tasian and Yong Fan},
  title = {{Fully-automatic segmentation of kidneys in clinical ultrasound images using a boundary distance regression network}},
  booktitle = {ISBI},
  year = {2019},
  url = {https://openreview.net/forum?id=CpXBQwlDSbey},
}
Figure