Chest X-ray abnormalities localization via ensemble of deep convolutional neural networks

Abstract

Convolutional neural networks have been applied widely in chest X-ray interpretation thanks to the availability of high-quality datasets. Among them, VinDr-CXR is one of the latest public datasets including 18000 expert-annotated images labeled into 22 local position-specific abnormalities and 6 globally suspected diseases. A proposed deep learning algorithm based on Faster-RCNN, Yolov5 and EfficientDet frameworks were developed and investigated in the task of multi-class clinical detection from chest radiography. The ground truth was defined by radiologist-adjudicated image review. Their performance was evaluated by the mean average precision. The results show that the best performance belonged to object detection models ensembled with an EfficientNet classifier, resulting in a peak mAP of 0.292. As a trade-off, ensembling object detection models was much slower, increasing computing time by 3.75, 5 and 2.25 times compared to FasterRCNN, Yolov5 and EfficientDet individually. Overall, the classifiers show constant improvement on all detector models, which is recommended for further research. All of these aspects should be considered to address real-world CXR diagnosis where the accuracy and computing cost are of concern.