Enhancing breast ultrasound segmentation through fine-tuning and optimization techniques: Sharp attention UNet

Area of Science:

• Medical Imaging
• Computer-Aided Diagnosis
• Deep Learning

Background:

• Accurate breast ultrasound image segmentation is vital for computer-aided diagnosis.
• Deep learning models, particularly UNet architectures, are powerful tools for medical image segmentation.
• Optimizing and fine-tuning these models is crucial for enhanced performance.

Purpose of the Study:

• To evaluate the impact of image preprocessing and optimization techniques on UNet models for breast ultrasound image segmentation.
• To compare the performance of standard UNet, Sharp UNet, and Attention UNet.
• To introduce and validate a novel hybrid model, Sharp Attention UNet.

Main Methods:

• Comparative analysis of image preprocessing and optimization techniques.
• Application of optimization and fine-tuning to UNet, Sharp UNet, and Attention UNet.
• Development and evaluation of the proposed Sharp Attention UNet model.

Main Results:

• The Sharp Attention UNet achieved a Dice coefficient of 0.93, specificity of 0.99, sensitivity of 0.94, and F1 score of 0.94.
• McNemar's statistical test indicated significant performance improvements.
• The proposed model outperformed existing approaches in breast lesion segmentation.

Conclusions:

• The Sharp Attention UNet model demonstrates superior performance in breast ultrasound image segmentation.
• Image preprocessing and fine-tuning are critical for optimizing deep learning models in this domain.
• The developed model offers a promising advancement for computer-aided diagnosis systems in breast imaging.