Enhancing Breast Ultrasound Segmentation through Fine-tuning and Optimization Techniques: Sharp Attention UNet.

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Deep Learning

Background:

• Breast ultrasound image segmentation is critical for computer-aided diagnosis but remains challenging.
• Deep learning (DL) models, particularly UNet architectures, are revolutionizing medical image analysis.
• Optimizing and fine-tuning DL models are essential for improving segmentation accuracy with increasing data complexity.

Approach:

• Evaluated the impact of image preprocessing, optimization, and fine-tuning techniques on UNet, Sharp UNet, and Attention UNet models.
• Developed a novel 'Sharp Attention UNet' by combining Sharp UNet and Attention UNet architectures.
• Applied McNemar's statistical test to compare the performance of different segmentation models.

Key Points:

• The proposed Sharp Attention UNet achieved a Dice coefficient of 0.9283, specificity of 0.9936, sensitivity of 0.9426, and F1 score of 0.9412.
• Demonstrated significant performance improvements over existing models through quantitative evaluation.
• Highlighted the importance of optimization and fine-tuning for enhanced UNet-based segmentation.

Conclusions:

• The Sharp Attention UNet model offers superior performance for breast lesion segmentation in ultrasound images.
• The findings suggest a pathway towards more accurate and reliable breast cancer diagnosis through advanced DL techniques.
• This research contributes to the advancement of automated medical image analysis for improved patient outcomes.