Aprendizado federado: Estudo da influência de artefatos sobre o modelo central na análise de microcalcificações por mamografia

Abstract

Background: The growing in use of machine learning techniques for medical diagnosis raises questions about the security of sensitive patient data and about sharing data for training, encouraging the study of innovative solutions to such problems. The federated learning technique allows the sharing of information for algorithm training in multicentric collaboration, with aggregation in a central server, without sharing databases, protecting sensitive data and eliminating the need for a large data stream for transmission. However, there is a lack of knowledge on the part of the central server regarding the composition of the databases used for training. When using convolutional networks for image diagnosis, the server's lack of knowledge concerns in particular the quality of the images used and the presence of artifacts. Objectives: Evaluate the influence of image artifacts in the training of a machine learning algorithm using the federated learning technique. Methods: Using mammography images, motion artifacts and dust artifacts were simulated, evaluating different training scenarios. An open source convolutional neural network was adapted to classify the images into 03 categories: images without microcalcifications; images with benign microcalcifications; and images with suspicious microcalcifications. Training was performed before and after the addition of simulated artifacts. Results: The addition of artifacts to images in the training databases influenced the performance of the algorithm, varying according to the size of the database and the type of artifact, with potential beneficial effects. Conclusion: The use of images with artifacts in the composition of databases can benefit the training of machine learning algorithms using a federated learning technique.