Related Experiment Video
Updated: Jun 13, 2025
04:48
Application of Deep Learning-Based Medical Image Segmentation via Orbital Computed Tomography
Published on: November 30, 2022
2.7K
Unsupervised non-small cell lung cancer tumor segmentation using cycled generative adversarial network with
Chengyijue Fang1, Xiaoyang Li2, Yidong Yang1,2
1Department of Engineering and Applied Physics, University of Science and Technology of China, Hefei, Anhui, China.
Journal of Applied Clinical Medical Physics
|April 23, 2025
Summary
This study introduces smic-GAN, an unsupervised deep learning method for lung tumor segmentation that matches supervised methods without manual annotations. This reduces data preparation workload and aids future supervised network training.
Area of Science:
- Medical Imaging
- Artificial Intelligence
- Oncology
Background:
- Accurate tumor segmentation is vital for lung disease diagnosis and treatment.
- Current deep learning methods often require extensive manual annotations, increasing preparation workload.
Purpose of the Study:
- To develop an unsupervised tumor segmentation network, smic-GAN, using a similarity-driven generative adversarial network with a cycle strategy.
- To eliminate the need for manual annotations, thereby reducing data preparation efforts.
Main Methods:
- Developed and trained smic-GAN on 504 lung cancer CT scans to generate tumor-free synthetic images.
- Obtained residual images by subtracting synthetic from original CT slices.
- Applied thresholding, 3D median filtering, and morphological operations to create binary tumor masks.
Main Results:
- smic-GAN demonstrated performance comparable to supervised methods (UNet, Incre-MRRN) and outperformed unsupervised cycle-GAN.
- Achieved a Dice similarity of 74.5% ± 11.2%, significantly better than cycle-GAN (69.1% ± 16.0%).
- Reported positive predictive values (PPV) and Hausdorff distances (HD95) competitive with supervised approaches.
Conclusions:
- smic-GAN offers comparable performance to supervised methods for lung tumor segmentation.
- The unsupervised nature significantly reduces the manual annotation workload for training data preparation.
- The method can serve as a valuable initial step for manual annotation in supervised network training.
