Lung lesions detection from CT images based on the modified Faster R-CNN

At the end of 2019, the outbreak of the COVID-19 epidemic brought huge economic losses all around the world, and also seriously affect human work, life and study. A large number of infected patients brought huge workload to doctors, but deep learning methods can effectively assist their diagnosis. This paper is based on Faster R-CNN, an end-to-end target detection model, to realize the detection of lesions in CT images of the novel coronavirus, which contributes to track the later condition of the confirmed patients and conduct timely treatment. Firstly, Kmeans++ was carried out to cluster the dimensions of the bounding box of the ground truth in annotated CT images, and appropriate anchors sizes and ratios were selected. Then, the performance of the Faster R-CNN model based on VGG-16 and ResNet-50 on the original datasets and the augment datasets is compared. Finally, the results show that, in the enhanced dataset, the Faster R-CNN model based on VGG-16 achieved a better performance, the Recall and Precision of which on the overall test set reached 68.12% and 65.58% respectively, and the missed detection rate(MR) was 31.88%.