Abstract:
The paper presents the results of the application of transfer learning
of deep convolutional neural networks for the task of searching for chest X-rays with
errors of patient styling and positioning. Evaluated neural network architectures:
InceptionV3, Xception, ResNet152V2, InceptionResnetV2, DenseNet201, VGG16,
VGG19, MobileNetV2, NASNetLarge. For training and testing we used chest X-rays
from open datasets and the unified radiological information service of the city of Moscow.
All the models obtained had diagnostic accuracy metrics above 95
based on the ResNet152V2, DenseNet201, VGG16, MobileNetV2 architectures had
statistically significantly better metrics than other models. The best absolute values
of metrics were shown by the ResNet152V2 model (AUC =0.999 , sensitivity=0.987,
specificity=0.988, accuracy=0.988, F1 score = 0.988). The MobileNetV2 model showed
the best processing speed of one study (67.8 $\pm$ 5.0 ms). The widespread use of the
algorithms we have obtained can facilitate the creation of large databases of high-quality
medical images, as well as optimize quality control when performing chest X-ray
examinations. (In Russian).
Key words and phrases:neural networks, deep learning, quality control, chest X-ray.