An Automatic Localization Algorithm for Ultrasound Breast Tumors Based on Human Visual Mechanism

Yuting Xie, Ke Chen, Jiangli LinSensors 2017 · DOI

TL;DR

This paper introduces a novel algorithm for automatically locating breast tumors in ultrasound images by adapting principles from the human visual system. By combining intensity, blackness ratio, and a new superpixel contrast feature, the method creates a saliency map to pinpoint the most likely tumor location, achieving a high accuracy of 94% on a dataset of 400 images.

What Problem Does This Paper Address?

Ultrasound (US) imaging is a crucial, non-invasive tool for breast cancer detection. However, US images often suffer from low contrast, speckle noise, and low spatial resolution, making it challenging for clinicians to accurately identify and delineate suspicious lesions. This difficulty can lead to misdiagnoses and increased workload for radiologists. While computer-aided diagnosis (CAD) systems aim to assist in this process, accurately localizing tumors in US images remains a significant hurdle, often requiring manual intervention. Existing automated methods, whether point-based or region-based, have limitations in handling noise, artifacts, and variations in tumor size and background complexity.

Methods

The core of this work is an adaptation of the Human Visual Mechanism (HVM) model, specifically Itti's model, to the domain of ultrasound breast imaging. The authors observed that tumors in US images tend to have higher global and local contrast compared to surrounding tissues. Their proposed method leverages this by computing a saliency map based on three key features:

  • Intensity: Tumors typically appear darker than surrounding tissues. This feature captures the basic intensity differences.
  • Blackness Ratio: To further emphasize darker regions and mitigate the influence of bright artifacts, a "blackness ratio" is calculated. This feature amplifies the contrast of low-intensity areas.
  • Superpixel Contrast: To address limitations of intensity and blackness ratio alone, especially in complex backgrounds or with speckle noise, a novel "superpixel contrast" feature is introduced. This involves segmenting the image into superpixels (regions of similar color and brightness) and then analyzing the contrast between these superpixels. This helps to capture more global and structural information.

These three features are processed through a Gaussian pyramid to capture information at different scales. Each feature generates its own saliency map, which are then combined. A "Winner Take All" (WTA) algorithm is employed on the combined saliency map to identify the most salient region, which is then localized as a potential tumor. A post-processing step, incorporating an "Inhibition of Return" (IOR) mechanism, is also used to refine localization and handle cases where the initial "winner" might be a false positive.

Datasets

The study was conducted using 400 ultrasound breast images obtained from the Ultrasound Department of West China Hospital of Sichuan University. The paper does not mention the creation of a new benchmark dataset, but rather uses this collection for evaluation.

Key Insights and Ideas

  • HVM Adaptation: The successful adaptation of a general visual attention model (HVM) to a specialized medical imaging domain (ultrasound breast tumors) highlights the potential of bio-inspired approaches in medical image analysis.
  • Multi-feature Fusion: Combining intensity, blackness ratio, and the novel superpixel contrast feature is crucial. Relying on a single feature is insufficient due to the inherent noise and artifacts in US images.
  • Superpixel Saliency: The introduction of superpixel contrast as a feature is a key innovation. It provides a more robust measure of local and global contrast, helping to distinguish tumors from complex backgrounds and artifacts that might confuse simpler intensity-based methods.
  • Addressing False Positives: The post-processing step with Inhibition of Return is important for improving accuracy by re-evaluating the most salient regions and potentially selecting a secondary candidate if the primary one is deemed incorrect.

Key Results

The proposed automatic localization algorithm achieved a high accuracy of 94.00% on a dataset of 400 ultrasound breast images, successfully localizing tumors in 376 of them. The method demonstrated effectiveness across a range of tumor sizes (small, medium, and large) and backgrounds, outperforming a compared method (Liu's method) which achieved 81.25% accuracy. The authors specifically noted that their method is better at handling images with serious speckle noise, artifacts, and larger tumors compared to the benchmark.

Limitations and Future Work

The authors acknowledge that the localization method has weaknesses, particularly in cases with complex background areas where it may struggle to differentiate tumors from other salient regions. They suggest that future work will focus on further improvements to reduce the interference of backgrounds on tumor localization. Additionally, they plan to extend the proposed method to other human tissues, such as the thyroid, kidney, and gallbladder.

Summary updated

Abstract

Human visual mechanisms (HVMs) can quickly localize the most salient object in natural images, but it is ineffective at localizing tumors in ultrasound breast images. In this paper, we research the characteristics of tumors, develop a classic HVM and propose a novel auto-localization method. Comparing to surrounding areas, tumors have higher global and local contrast. In this method, intensity, blackness ratio and superpixel contrast features are combined to compute a saliency map, in which a Winner Take All algorithm is used to localize the most salient region, which is represented by a circle. The results show that the proposed method can successfully avoid the interference caused by background areas of low echo and high intensity. The method has been tested on 400 ultrasound breast images, among which 376 images succeed in localization. This means this method has a high accuracy of 94.00%, indicating its good performance in real-life applications.
BibTeX
@inproceedings{xie2017automatic,
  author = {Yuting Xie and Ke Chen and Jiangli Lin},
  title = {{An Automatic Localization Algorithm for Ultrasound Breast Tumors Based on Human Visual Mechanism}},
  booktitle = {Sensors},
  year = {2017},
  doi = {10.3390/s17051101},
}
Figure