Automatic Identification of Aspiration Pneumonia Based on Bronchoscope Images and Deep Learning
Dawei Gong,
Ke Cui,
Weidong Wang,
Xiaobo Chen,
Chao Zhang,
Haifei Xiang,
Shaohua Zhang and
Sailing He
Aspiration pneumonia is a type of lung infection caused by the accidental inhalation of foreign substances into the respiratory tract. It is commonly seen in the elderly, young children, and individuals who are unconscious or have difficulty swallowing. Early detection and diagnosis of aspiration pneumonia are beneficial for improving patient outcomes and reducing the medical burden. In this study, we collected bronchoscopic video data from 25 patients in two hospitals. After image preprocessing and expert annotation, we obtained 2830 images from some patients for training and 1215 images from the other patients for validation. We selected three deep learning methods for training. The experimental test results for the identification of aspiration pneumonia showed that ResNet-50, which is based on convolutional operations, gave the best performance in the automatic identification of aspiration pneumonia, with a precision of 97.82%, a recall of 91.82%, an F1 score of 94.73%, and an overall accuracy of 95.88%. The experiments demonstrated that deep learning methods can be used for the automatic identification and diagnosis of aspiration pneumonia from bronchoscope images and deep learning is reported here for the first time for diagnosing aspiration pneumonia from bronchoscope images.
