基于深度学习的肺炎图像目标检测

肺炎是一种严重危害身体健康的疾病,通常使用肺部X光片进行检查。肺炎诊断是肺炎治疗前非常重要的环节,但是由于肺部其他疾病的干扰、医疗数据的爆发式增长以及专业病理医生的缺乏等,导致肺炎的准确诊断较为困难。深度学习能够模仿人脑的机制准确高效地解释医学图像数据,在肺炎图像检测方面获得了广泛应用。构建了3种基于深度学习的图像目标检测模型,单发多框探测器(SSD)、faster-RCNN和faster-RCNN优化模型,对来自Kaggle数据集的26 684张带标签的肺部X光图像进行研究。原始X光图像经预处理后输入3种深度学习模型,分别对单处和两处病灶区域进行目标检测。随机选取500张测试图像,利用损失函数、分类准确率、回归精度和误检病灶数等指标对各模型的性能进行评估。结果表明,faster-RCNN的性能指标优于SSD;Faster-RCNN优化模型的性能指标均优于其他两种模型,其损失函数值小且可快速达到稳定,平均分类准确率为93.7%,平均回归精度为79.8%,且误检病灶数为0。该方法有助于肺炎的准确识别和诊断。

Object Detection of Pneumonia Images Based on Deep Learning

Pneumonia is a disease that seriously endangers people′s health. Lung X-rays are usually used for pneumonia examination. The diagnosis of pneumonia is a very important step before the treatment of pneumonia. However, due to the interference of other lung diseases, the explosion of medical data, and the lack of professional pathologists, it is very difficult to accurately diagnose pneumonia. Deep learning can imitate the mechanism of the human brain to interpret medical image datasets with improved accuracy and efficiency, therefore, has been widely used in pneumonia image detection. In this paper, three deep learning-based object detection models, SSD, faster-RCNN and faster-RCNN optimization model, were used to study 26 684 labeled lung X-ray images from the Kaggle dataset. The original X-ray images were preprocessed and then input into the three deep learning models to detect single or two lesion areas. The performance of the three models was evaluated and compared through loss function, classification accuracy, regression accuracy and number of mis-detected lesions by testing 500 randomly selected images. The results showed that faster-RCNN performed better than SSD in performance metrics; Faster-RCNN optimization model was superior to the other two models with the loss value that was small and could quickly reach stability, the average classification accuracy of 93.7%, the average regression accuracy of 79.8% and the number of mis-detected lesions of 0, which would be helpful for the accurate identification and diagnosis of pneumonia.