Related Experiment Video
Updated: Dec 26, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
3.3K
Lesion Segmentation in Ultrasound Using Semi-Pixel-Wise Cycle Generative Adversarial Nets
Summary
A new deep learning method, semi-pixel-wise cycle generative adversarial net (SPCGAN), accurately segments breast cancer lesions in ultrasound images. This approach is particularly effective for malignant lesions and when training data is limited.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Oncology
Background:
- Breast cancer is the most common cancer in females, necessitating accurate diagnostic tools.
- Handheld ultrasound is a key imaging modality for breast cancer identification.
- Lesion segmentation in ultrasound images aids clinical decision-making.
Purpose of the Study:
- To propose a novel deep learning method, semi-pixel-wise cycle generative adversarial net (SPCGAN), for segmenting breast lesions in 2D ultrasound.
- To evaluate the performance of SPCGAN against existing segmentation techniques.
Main Methods:
- Developed a semi-pixel-wise cycle generative adversarial net (SPCGAN) utilizing a fully convolutional neural network (FCN) and generative adversarial network.
- Compared SPCGAN with a fully connected neural network and level set segmentation on a dataset of 32 malignant and 109 benign lesions.
Main Results:
- SPCGAN achieved a Dice similarity coefficient (DSC) of 0.92, outperforming FCN (0.90) and level set (0.79).
- For malignant lesions, SPCGAN significantly improved DSC from 0.90 to 0.93 (p < 0.001).
- SPCGAN demonstrated robust segmentation and effectiveness with limited training data.
Conclusions:
- The proposed SPCGAN method provides robust and accurate segmentation of breast lesions in ultrasound images.
- SPCGAN is particularly advantageous in scenarios with limited training data, potentially reducing radiologists' annotation burden.
