Related Experiment Video
Updated: Sep 3, 2025
04:23
A Swin Transformer-Based Model for Thyroid Nodule Detection in Ultrasound Images
Published on: April 21, 2023
1.9K
A Comparative Study of Multiple Deep Learning Models Based on Multi-Input Resolution for Breast Ultrasound Images
Huaiyu Wu1, Xiuqin Ye1, Yitao Jiang2,3
1Department of Ultrasound, First Clinical College of Jinan University, Second Clinical College of Jinan University, First Affiliated Hospital of Southern University of Science and Technology, Shenzhen People's Hospital, Shenzhen, China.
Frontiers in Oncology
|July 25, 2022
Summary
This study evaluated deep learning (DL) models for breast cancer diagnosis using ultrasound images. The best DL model combinations outperformed physicians in accuracy, demonstrating AI
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Oncology
Background:
- Breast cancer diagnosis relies heavily on medical imaging interpretation.
- Deep learning (DL) models offer potential for improving diagnostic accuracy and efficiency.
- Optimizing DL model parameters and input image resolutions is crucial for clinical application.
Purpose of the Study:
- To evaluate the performance of various deep learning models (Xception, DenseNet121, MobileNet, ResNet50, EfficientNetB0) for breast cancer diagnosis.
- To investigate the impact of different input image resolutions (224x224, 320x320, 488x488 pixels) on DL model performance.
- To compare the diagnostic accuracy of optimized DL models against human physicians.
Main Methods:
- Retrospective analysis of 13,684 grayscale ultrasound breast images from two Chinese hospitals.
- Training, validation, and internal/external testing of five DL models with varying parameter combinations and resolutions.
- Comparison of DL model performance (AUC, sensitivity, specificity, accuracy) against physicians using Wilcoxon test.
Main Results:
- EfficientNetB0 with 320x320 resolution achieved an AUC of 0.907 in external testing.
- Xception with 448x448 resolution achieved an AUC of 0.900 in external testing.
- Optimized DL models demonstrated superior diagnostic performance compared to senior and junior physicians in physician-AI comparison tests.
Conclusions:
- Specific deep learning models combined with optimized image resolutions show high diagnostic performance for breast cancer.
- The best performing DL model combinations exceeded the diagnostic accuracy of physicians, indicating AI's potential in clinical settings.
- Further validation is warranted to integrate these AI tools into routine breast cancer screening and diagnosis.
