Image Generation and Lesion Segmentation of Brain Tumors and Stroke Based on GAN and 3D ResU-Net
In recent years, with the development of medical image analysis, the image processing of brain tumor and stroke has been deeply studied. However, when processing clinical medical imaging data with different characteristics or information collected from different sensors or modalities, that is, multi...
Saved in:
| Main Authors: | , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
IEEE
2025-01-01
|
| Series: | IEEE Access |
| Subjects: | |
| Online Access: | https://ieeexplore.ieee.org/document/10487941/ |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | In recent years, with the development of medical image analysis, the image processing of brain tumor and stroke has been deeply studied. However, when processing clinical medical imaging data with different characteristics or information collected from different sensors or modalities, that is, multi-modal imaging data, the segmentation accuracy is low. Therefore, the research is based on generative adversarial networks and three-dimensional residual U-shaped networks to study brain tumor and stroke image generation and lesion segmentation. Experimental results showed the three models performed best in various conversions. For example, in T1<inline-formula> <tex-math notation="LaTeX">$\to $ </tex-math></inline-formula> Flair conversion, the generative multi-modal image analysis model based on perceptual loop consistency had an average peak signal-to-noise ratio of <inline-formula> <tex-math notation="LaTeX">$23.951~\pm ~2.735$ </tex-math></inline-formula>, an average structural similarity of <inline-formula> <tex-math notation="LaTeX">$0.873~\pm ~0.046$ </tex-math></inline-formula>, and an average root mean square error of <inline-formula> <tex-math notation="LaTeX">$16.998~\pm ~6.184$ </tex-math></inline-formula>. All three models significantly raised the segmentation effectiveness of lesions, such as the combination of dual-scale perceptual loop generation adversarial network and three-dimensional residual U-shaped network for generative multi-modal image generation and lesion segmentation algorithm. Using three real input modalities, its HD index value of 75.082 and precision index value of 0.696 were better than the HD index value of 84.776 and precision index value of 0.686. In addition, the study also conducted ablation experiments on a generative multi-modal image analysis model based on dual-scale perceptual loop consistency, indicating that the cavity residual module is hoped to have a good influence on lesion segmentation. Overall, the algorithm model proposed in the study has high effectiveness in the generation and segmentation of brain tumors and stroke images, and is of great significance for the development of medicine. |
|---|---|
| ISSN: | 2169-3536 |