Transfer Learning for automated search for defects on chest X-rays

Abstract

Purpose. To develop and test algorithms for determining the projection and searching for common technical defects on chest -rays using transfer learning with various neural network architectures. Materials and methods. Algorithms have been created to search for technical remarks such as incorrect choice of study boundaries and errors of patient positioning. Transfer learning of neural network architectures VGG19 and ResNet152V2 was chosen as the basis for creating algorithms. To train and test the algorithms, we used radiographs from open databases (over 230,000 studies in total). To validate the obtained algorithms, a test dataset was prepared from 150 anonymized chest x-rays unloaded from the Unified Radiological Information Service of the Moscow city (URIS) and evaluated by expert doctors and technicians. Results. All obtained algorithms have high classification quality indicators. The maximum accuracy on the test dataset was obtained for the model that determines the projection, AUC was 1.0, the minimum accuracy: AUC 0.968 was obtained for the model that determines the rotation of the chest on the lateral X-ray. On the validation dataset maximum accuracy was obtained for the model that determines the projection, AUC was 0.996, the minimum accuracy: AUC 0.898 was obtained for the model that determines the rotation of the chest on the lateral x-ray. Conclusions. All of diagnostic accuracy metrics for each of the models exceeded the threshold value of 0.81 and can be recommended for practical use.

Ähnliche Arbeiten