对结核病检测进行深度学习模型的比较:对数字与模拟胸部放射图的回顾性研究
Subhankar Chattoraj1, Bhargava Reddy1, Manoj Tadepalli1
1Qure.ai, Mumbai, Maharashtra, India.
The Indian journal of tuberculosis
|December 8, 2025
概括
一个深度学习设备在胸部X射线上准确检测结核病 (TB),在数字和模拟图像上都可靠地执行. 这种自动化解决方案对资源有限的设置有希望,改善结核病诊断.
科学领域:
- 放射学 放射学是一门学科.
- 人工智能的人工智能
- 医疗成像医学成像
背景情况:
- 胸部X射线 (CXR) 对于检测肺部异常至关重要,每年在全球范围内进行超过20亿次.
- 准确及时诊断结核病 (TB) 仍然是一个挑战,特别是在中低收入国家.
- 需要便携式,自动化和可靠的结核病检测解决方案.
研究的目的:
- 为了比较基于深度学习 (DL) 的设备在数字和模拟CXR上进行结核病检测的性能.
- 为了评估DL设备在资源受限制的设置中的实用性.
主要方法:
- 一万张CXR图像 (来自手机的DICOM和打印照片) 被追溯收集.
- 基于DL的设备的性能使用这些CXR及其相关的放射学报告进行了评估.
- 图像是使用三星S8,iPhone 8和iPhone XS获得的.
主要成果:
- 该DL装置在DICOM上实现了0.928的曲线下面面积 (AUC),具有0.841的灵敏度和0.806的TB检测特异性.
- 在将DL设备在手机图像 (模拟CXR) 和原始DICOMs上的性能进行比较时,观察到AUC的最小差异 (≤5.10%).
- 这些发现证明了DL设备在数字和模拟CXR格式的稳定性.
结论:
- 基于DL的设备在识别数字和模拟CXR的结核病放射性迹象方面表现出强大的性能.
- 在不同的图像类型中保持一致的性能突出显示了该设备对结核病检测的可靠性.
- 该设备显示出临床实用性的潜力,特别是在资源有限的环境中,需要自动结核病查.
相关概念视频
Pulmonary Tuberculosis IV
468
Tuberculosis, more commonly referred to as TB, is an infectious disease stemming from Mycobacterium tuberculosis. While it primarily impacts the lungs, TB can also affect other body areas. Given its severity and global impact, timely and accurate diagnosis is crucial for controlling its spread and improving patient outcomes.
Several diagnostic approaches are used to detect TB. The conventional method is the Tuberculin Skin Test (TST), also known as the Mantoux test. However, this method has...
Several diagnostic approaches are used to detect TB. The conventional method is the Tuberculin Skin Test (TST), also known as the Mantoux test. However, this method has...
468
Pulmonary Tuberculosis III
899
Tuberculosis (TB) is a contagious infection primarily affecting the lung parenchyma but which can also affect other body parts. TB can be classified based on disease development, presentation, and the affected anatomical site.
The first classification is based on the development of the disease, and it includes the following categories:
The first classification is based on the development of the disease, and it includes the following categories:
899


