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

Shi Yin, Zhengqiang Zhang+6 more|arXiv (Cornell University)|Jan 5, 2019

Advanced Neural Network Applications20 references4 citations

TL;DR

This paper proposes a fully automatic kidney segmentation method in clinical ultrasound images using a boundary distance regression network. By leveraging pre-trained features and learning distance maps to the kidney boundaries, the approach improves segmentation accuracy over standard pixel classification networks, achieving state-of-the-art performance on challenging ultrasound data with variable anatomy and intensity distributions.

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.

Motivation & Objective

• To address the challenge of automatic kidney segmentation in clinical ultrasound due to variable kidney shapes and intensity distributions.
• To develop a fully automatic method that reduces reliance on manual or semi-automatic intervention.
• To improve segmentation accuracy by modeling boundary consistency across diverse ultrasound images.
• To integrate deep feature extraction with boundary-aware regression for robust organ delineation.

Proposed method

• Utilizes pre-trained deep neural networks from natural image classification to extract high-level features from ultrasound images.
• Employs a boundary distance regression network to predict distance maps from each pixel to the nearest kidney boundary.
• Applies an end-to-end learning framework where the predicted distance maps are used to classify pixels as kidney or non-kidney.
• Uses a pixel classification network that leverages the boundary distance maps for final segmentation prediction.
• Trains the entire system in an end-to-end fashion to optimize segmentation performance.
• Leverages the geometric consistency of kidney boundaries across patients to improve generalization and robustness.

Experimental results

Research questions

• RQ1 Can boundary distance regression improve automatic kidney segmentation in ultrasound compared to standard pixel classification?
• RQ2 How does using pre-trained features from natural images enhance performance on medical ultrasound segmentation?
• RQ3 To what extent does modeling boundary consistency improve segmentation accuracy across diverse kidney shapes and image intensities?
• RQ4 Can an end-to-end learning framework combining feature extraction, distance regression, and classification outperform existing deep learning baselines?

Key findings

• The proposed method significantly outperforms standard deep learning-based pixel classification networks in kidney segmentation accuracy.
• The use of boundary distance maps leads to more precise localization of kidney boundaries, especially in low-contrast or noisy ultrasound images.
• The integration of pre-trained features from natural images improves feature representation for ultrasound, enhancing generalization.
• The end-to-end training strategy effectively optimizes the full pipeline, resulting in robust segmentation performance.
• The method demonstrates strong generalization across diverse patient populations and scanning protocols.
• The approach achieves state-of-the-art performance on the evaluated dataset, confirming the effectiveness of boundary-aware regression in medical image segmentation.