构建学习模型在痰液抗酸染色中识别结核分枝杆菌的应用

The application of deep learning in the identification of Mycobacterium tuberculosis in acid⁃fast staining of sputum

目的构建能准确识别抗酸染色痰涂片中的结核分枝杆菌(Mtb)的学习模型,为更快、更高效、更准确地诊断肺结核提供技术支持。方法从PubMed数据库已发表的文章和Tuberculosis Image Dataset中收集痰液Mtb抗酸染色阳性图像,通过直接将像素值除以255对所有图像进行归一化,并运用Batch Normalization进行标准化处理,利用迁移学习技术构建快速卷积神经网络(Faster R⁃CNN)深度学习目标检测模型。计算模型精确度、交并比(IoU)、均值平均精度(mAP),并绘制迭代次数-损失折线图对模型的性能进行评价。结果共收集Tuberculosis Image Dataset中的1265张Mtb抗酸染色图像和PubMed数据库中33篇文章中的60张Mtb抗酸染色图像,用于Faster R⁃CNN深度学习目标检测模型构建。该模型表现出较好的预测效果,迭代次数-损失折线图相对平滑,呈现出明显的下降趋势,约在迭代60次左趋于稳定状态。训练集的精确度、IoU、mAP分别为0.9489、0.8975、0.8543,验证集的精确度、IoU、mAP分别为0.6931、0.5460、0.8453。测试集共含有60张Mtb阳性图片,其中49张被模型正确预测,预测结果的正确率为81.7%。结论Faster R⁃CNN深度学习目标检测模型对痰液抗酸染色阳性的Mtb具有较好的识别和分类能力,可作为肺结核诊断的辅助工具用于临床环境。

Objective To establish a learning model that could accurately identify Mycobacterium tuberculosis(Mtb)in acid⁃fast stained sputum smears,and provide technical support for faster,more efficient and more accurate diagnosis of pulmonary tuberculosis.Methods The positive images of Mtb acid⁃fast staining in sputum were collected from the published literature of PubMed database and Tuberculosis Image Dataset.All images were normalized by directly dividing the pixel values by 255,and standardized using Batch Normalization.Faster R⁃CNN using transfer learning technology was used to construct a deep learning object detection model.The performance of the model was evaluated by calculating the model accuracy,intersection over union(IoU),and mean average precision(mAP),and drawing the loss⁃epochs curve.Results A total of 1265 Mtb acid⁃fast staining images from the Tuberculosis Image Dataset and 60 Mtb acid⁃fast staining images from 33 literatures in PubMed database were collected for the construction of Faster R⁃CNN deep learning object detection model.The model showed good prediction effect.The loss⁃epochs curve was relatively smooth,showing an obvious downward trend,and tended to be stable at about 60 iterations.The accuracy,IoU and mAP of the training set were 0.9489,0.8975 and 0.8543,respectively,and those of the validation set were 0.6931,0.5460 and 0.8453,respectively.The testing set comprises a total of 60 Mtb⁃positive images;out of which,the model accurately predicted 49,with an accuracy of 81.7%.Conclusion The Faster R⁃CNN deep learning objective detection model had a good ability to identify and classify acid⁃fast staining positive Mtb in sputum,which could be used as an auxiliary tool for clinical diagnosis of pulmonary tuberculosis.