Abstract：
Tuberculosis (TB) poses a significant health threat in Taiwan, necessitating efficient detection methods. Traditional screening for acid-fast positive bacilli in acid-fast stain is time-consuming and prone to human error due to staining artifacts. To address this, we present an automated TB detection platform leveraging deep learning and image processing. Whole slide images from 2 hospitals were collected and processed on a high-performance system. The system utilizes an image processing technique to highlight red, rod-like regions and a modified EfficientNet model for binary classification of TB-positive regions. Our approach achieves a 97% accuracy in tile-based TB image classification, with minimal loss during the image processing step. By setting a 0.99 threshold, false positives are significantly reduced, resulting in a 94% detection rate when assisting pathologists, compared with 68% without artificial intelligence assistance. Notably, our system efficiently identifies artifacts and contaminants, addressing challenges in digital slide interpretation. Cross-hospital validation demonstrates the system\'s adaptability. The proposed artificial intelligence-assisted pipeline improves both detection rates and time efficiency, making it a promising tool for routine pathology work in TB detection.
摘要：
结核病（TB）在台湾构成重大健康威胁，需要有效的检测方法。在耐酸染色剂中对耐酸阳性杆菌的传统筛选是耗时的,并且由于染色伪影而容易出现人为错误。为了解决这个问题，我们提出了一个利用深度学习和图像处理的自动结核病检测平台。在高性能系统上收集并处理来自2家医院的整个幻灯片图像。该系统利用图像处理技术突出显示红色，杆状区域和改进的EfficientNet模型，用于对TB阳性区域进行二元分类。我们的方法在基于图块的TB图像分类中实现了97%的准确率，在图像处理步骤中损失最小。通过设置0.99阈值，假阳性显著减少，在协助病理学家时，检出率为94%，与没有人工智能援助的68%相比。值得注意的是,我们的系统有效地识别了工件和污染物，解决数字幻灯片解释中的挑战。跨医院验证证明了系统的适应性。提出的人工智能辅助管道提高了检测率和时间效率，使其成为结核病检测中常规病理学工作的有前途的工具。
