在疾病领域提高超声波图像质量:应用循环一致的生成对抗网络和感知损失
Shreeram Athreya1, Ashwath Radhachandran2, Vedrana Ivezić3
1Department of Electrical and Computer Engineering, University of California Los Angeles, Los Angeles, CA, United States.
JMIR biomedical engineering
|December 17, 2024
概括
这项研究引入了一个增强的CycleGAN模型,具有感知损失,以改善便携式设备的低质量超声波图像. 该框架有效处理未注册的图像对,提高诊断质量和可访问性.
科学领域:
- 医疗成像医学成像
- 人工智能的人工智能
- 图像处理 图像处理
背景情况:
- 现有的超声波图像增强方法经常使用注册图像或器官特征,限制了概括性.
- 与高端设备相比,便携式超声波设备产生的图像质量较低,造成了性能差距.
- 需要一个通用的框架来提高超声波图像质量,在各种设备和条件下,独立于注册或器官特征.
研究的目的:
- 开发一个强大的框架来提高超声波图像质量,特别是从低成本的便携式设备.
- 在各种临床环境和设备类型中有效处理未注册的超声波图像对.
- 为改善诊断成像可访问性和质量提供一个通用和可适应的解决方案.
主要方法:
- 使用 CycleGAN 框架增强的知觉损失进行了回顾性分析.
- 通过比较预训练的神经网络的深度特征,整合了感知损失,以保持解剖完整性.
- 模型的性能与高分辨率图像进行了评估,并在多样化的公共数据集上进行了验证.
主要成果:
- 增强的CycleGAN框架在关键评估指标上显著优于稳定的CycleGAN.
- 结构相似性指数改善至0.2889 (相对于0.2502),峰值信号与噪声比为15.8935 (相对于14.9430).
- 学习感知图像补丁相似性得分为0.4490 (vs. 0.5005),表明优越的增强和解剖细节的保存.
结论:
- 增强的CycleGAN模型成功地弥合了不同设备之间的超声波成像质量差距.
- 通过处理未注册的图像对,该模型提高了视觉质量,同时保留了用于诊断的关键解剖结构.
- 这种方法可以使高质量的超声波成像实现民主化,改善医疗保健结果,特别是在资源有限的环境中.
相关概念视频
Ultrasonography
4.3K
Ultrasonography is an imaging technique that uses high-frequency sound waves to visualize the body's internal structures. It is a non-invasive and safe procedure that does not involve the use of ionizing radiation, making it widely used in various medical fields. Ultrasonography is used to study heart function, blood flow in the neck or extremities, certain conditions such as gallbladder disease, and fetal growth and development.
During an ultrasonography procedure, a handheld device called...
During an ultrasonography procedure, a handheld device called...
4.3K
Upsampling
202
Managing signal sampling rates is essential in digital signal processing to maintain signal integrity. A decimated signal, characterized by a reduced frequency range due to its lower sampling rate, can be upsampled by inserting zeros between each sample. This upsampling process expands the original spectrum and introduces repeated spectral replicas at intervals dictated by the new Nyquist frequency. To refine this zero-inserted sequence, it is passed through a lowpass filter with a cutoff...
202


