Related Experiment Video
Updated: Sep 8, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.9K
Enhancing Lesion Segmentation in Ultrasound Images: The Impact of Targeted Data Augmentation Strategies
Xu Wang1, Patrice Monkam1,2, Bonan Zhao1
1College of Medicine and Biological Information Engineering, Northeastern University, Shenyang, Liaoning, China.
International Journal of Biomedical Imaging
|August 20, 2025
Summary
This study introduces five data augmentation strategies to improve automated lesion segmentation in ultrasound images, especially when annotated data is scarce. These methods significantly enhance diagnostic accuracy for breast and thyroid lesions using deep learning models.
Area of Science:
- Medical imaging
- Artificial intelligence in healthcare
- Computer-assisted diagnosis
Background:
- Automated lesion segmentation in ultrasound (US) images using deep learning (DL) is vital for disease diagnosis.
- Limited annotated datasets hinder the practical application of DL models, particularly in resource-constrained settings.
- Generative adversarial networks (GANs) offer solutions for data scarcity but involve complex training and high computational costs.
Purpose of the Study:
- To explore novel solutions for automated lesion delineation in US images, addressing the challenge of limited annotated samples.
- To propose and evaluate five distinct mixed sample augmentation strategies for US lesion segmentation.
- To assess the effectiveness of these strategies across different lesion types (breast, thyroid) and deep segmentation models.
Main Methods:
- Development and implementation of five mixed sample data augmentation strategies.
- Evaluation of these strategies using four deep segmentation models.
- Performance assessment for breast and thyroid lesion delineation using Dice and Jaccard indices.
Main Results:
- Augmentation strategies significantly improved performance, with Dice and Jaccard indices increasing by up to 37.95% and 36.32% for breast lesions, and 14.59% and 13.01% for thyroid lesions.
- The effectiveness of augmentation strategies varied depending on lesion type and model architecture.
- Strategic selection of data augmentation approaches proved critical for enhancing DL model performance.
Conclusions:
- The proposed mixed sample augmentation strategies offer a reliable solution to data scarcity in automated US lesion segmentation.
- Careful selection of data augmentation techniques is crucial for maximizing performance gains in DL models for medical imaging.
- These findings provide valuable insights for improving automated diagnostic tools in healthcare.
