Breast ultrasound lesion classification based on image decomposition and transfer learning

Area of Science:

• Medical Imaging
• Deep Learning
• Biomedical Engineering

Background:

• Deep learning shows promise in medical image analysis, particularly for breast lesion classification using ultrasound.
• Acquiring large ultrasound datasets for breast lesions is challenging, often necessitating transfer learning.
• Traditional convolutional neural networks struggle with feature extraction from noisy ultrasound images, leading to unreliable classifications.

Purpose of the Study:

• To develop an improved method for extracting valuable information from breast lesion ultrasound images.
• To enhance the accuracy of benign and malignant breast lesion classification.
• To overcome limitations of traditional transfer learning in medical image analysis.

Main Methods:

• Image decomposition techniques, including fuzzy enhancement and bilateral filtering, were used to create multifeature data from original ultrasound images.
• A pre-trained VGG16 model was employed for knowledge fusion and feature extraction from the combined multifeature dataset.
• Fully connected layers were trained using the extracted features and expert-provided ground truths for classification.

Main Results:

• The multifeature approach achieved high classification performance: 93% accuracy, 95% sensitivity, 88% specificity, an F1 score of 0.93, and an AUC of 0.97.
• Feature extraction from multifeature data using a pre-trained VGG16 model proved effective for breast lesion classification.
• The fused features enabled accurate training of fully connected layers for reliable classification.

Conclusions:

• Multifeature data derived from image decomposition significantly improves feature extraction compared to using single ultrasound images.
• This enhanced feature extraction leads to superior breast lesion classification results.
• The proposed method offers a more reliable alternative to traditional transfer learning techniques for medical image analysis.