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

Area of Science:

• Medical imaging
• Computer vision
• Biomedical engineering

Background:

• Standard human visual mechanisms (HVMs) effectively identify salient objects in natural images but struggle with tumor localization in ultrasound breast images.
• Tumors in ultrasound images exhibit distinct characteristics, including higher global and local contrast compared to surrounding tissues.

Purpose of the Study:

• To research the specific characteristics of tumors in ultrasound breast images.
• To develop a novel automated method for accurate tumor localization.
• To overcome the limitations of existing human visual mechanisms in this diagnostic context.

Main Methods:

• A classic human visual mechanism (HVM) was adapted and enhanced.
• A novel auto-localization method was developed, integrating intensity, blackness ratio, and superpixel contrast features.
• A saliency map was computed, and a Winner Take All algorithm identified the most salient region, represented as a circle.

Main Results:

• The proposed method successfully distinguished tumors from background noise, including areas of low echo and high intensity.
• Tested on 400 ultrasound breast images, the method achieved successful localization in 376 cases.
• An accuracy rate of 94.00% was demonstrated, indicating robust performance.

Conclusions:

• The novel auto-localization method demonstrates high accuracy and effectiveness for tumor detection in ultrasound breast imaging.
• This approach offers a significant improvement over traditional human visual mechanisms for this specific medical imaging task.
• The method shows strong potential for real-world clinical applications in breast cancer screening and diagnosis.