Breast ultrasound image classification based on multiple-instance learning.

Area of Science:

• Medical Imaging
• Machine Learning
• Computer-Aided Diagnosis

Background:

• Breast ultrasound (BUS) image segmentation is challenging due to poor quality and speckle noise.
• Accurate breast tumor classification is crucial for diagnosis and treatment planning.
• Traditional multiple-instance learning (MIL) methods struggle with the complexity of BUS images.

Purpose of the Study:

• To develop a novel MIL method for improved breast tumor classification using BUS images.
• To address the limitations of traditional MIL in handling noisy and complex BUS data.
• To enhance the accuracy and reliability of computer-aided diagnosis for breast cancer.

Main Methods:

• Utilized local features from roughly segmented regions of interest (ROIs) as instances within a 'bag'.
• Employed a self-organizing map to map instance space to concept space.
• Constructed bag feature vectors based on instance distribution in the concept space.
• Applied a support vector machine (SVM) for final tumor classification.

Main Results:

• Achieved a classification accuracy of 0.9107.
• Obtained an area under the receiver operator characteristic curve (AUC) of 0.96 (p < 0.005).
• Demonstrated superior performance compared to traditional MIL methods for BUS image analysis.

Conclusions:

• The proposed novel MIL method effectively classifies breast tumors in BUS images.
• The approach successfully handles the inherent complexities and noise in BUS data.
• This method offers a promising advancement for computer-aided diagnosis in breast cancer screening.