GSDA: Generative adversarial network-based semi-supervised data augmentation for ultrasound image classification.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Computer Vision

Background:

• Medical Ultrasound (US) is a vital imaging tool, but variable image quality poses challenges.
• Deep Learning (DL) models offer advanced analysis but require extensive datasets, which are often scarce in medical US.
• Existing DL models struggle with limited data, hindering their performance in medical US image analysis.

Purpose of the Study:

• To develop a semi-supervised data augmentation method, GSDA, to address the data scarcity in medical US.
• To leverage Generative Adversarial Networks (GANs) for synthesizing high-resolution, high-quality US images.
• To improve the performance of Convolutional Neural Networks (CNNs) for medical US image analysis using augmented data.

Main Methods:

• Developed GSDA, a GAN-based semi-supervised data augmentation method combining GANs and CNNs.
• Employed transfer learning techniques to overcome training challenges with limited data for both GAN and CNN.
• Introduced a novel evaluation standard balancing classification accuracy and computational time.

Main Results:

• GSDA successfully synthesized high-resolution, high-quality pseudo-labeled US images.
• The method achieved a 97.9% accuracy on the BUSI dataset using only 780 images.
• GSDA demonstrated superior performance compared to existing state-of-the-art methods.

Conclusions:

• GSDA effectively addresses the data shortage challenge in medical US imaging through data augmentation.
• The synthesized high-quality images significantly enhance DL model performance.
• GSDA shows strong potential as an auxiliary tool for improving medical US analysis.