A cGAN-based tumor segmentation method for breast ultrasound images

Area of Science:

• Medical Imaging
• Artificial Intelligence in Medicine
• Biomedical Engineering

Background:

• Accurate segmentation of breast tumors in ultrasound images is crucial for diagnosis and treatment planning.
• Existing methods often struggle with image quality and precise tumor boundary delineation.
• Data augmentation techniques are needed to improve the robustness of segmentation models.

Purpose of the Study:

• To develop a novel method for enhancing breast ultrasound images and segmenting tumors.
• To improve the realism of enhanced ultrasound images and the accuracy of tumor segmentation.
• To introduce an optimized EfficientUNet architecture for precise medical image segmentation.

Main Methods:

• Utilized a conditional generative adversarial network (cGAN) for data enhancement and dataset expansion.
• Employed adversarial training to improve spatial consistency, replacing traditional post-processing methods.
• Proposed an EfficientUNet model integrating ResNet18, an attention mechanism (SENet), and deep supervision for enhanced feature extraction and faster convergence.

Main Results:

• The proposed cGAN method effectively enhances the quality and realism of breast ultrasound images.
• The optimized EfficientUNet achieved high segmentation performance, with Dice and IoU scores of 0.8856 and 0.8111, respectively, on Dataset B.
• The integration of ResNet18 and SENet improved feature extraction and tumor boundary detection.

Conclusions:

• The combined approach of cGAN for data enhancement and optimized EfficientUNet for segmentation offers a significant advancement in breast tumor analysis.
• The study demonstrates the efficacy of generative adversarial training and advanced neural network architectures in medical image analysis.
• This work provides a more accurate and robust solution for breast tumor segmentation from ultrasound images.