Evaluating the Inclusion of Images with Artifacts in Medical Image Databases of Mammography for Machine Learning
HeráClio Almeida Da Costa, Edmar Candeia Gurjão, Vitor Trindade Rocha Ribeiro · 2023
The use of the federated learning technique has grown in the medical field of diagnostic imaging, with a large number of publications and applications in clinical practice. However, the lack of knowledge of the training data by the central server and the clients among themselves raises concerns about the local composition of the databases, raising questions about the occurrence of image artifacts and their impact on the final performance of the global model. Using databases with a selection of areas of interest for mammograms distributed in an unbalanced way to 03 clients, simulations of mammography image artifacts were proceeded, with their insertion in the databases on different scenarios. An open access convolutional neural network model (EfficientNetB7) was adapted to classify the images regarding the presence and suspicion of microcalcifications according to the BI-RADS system. By measuring the F1-score, the performance of the neural network was verified without the presence of artifacts, after the insertion of artifacts and after the exclusion of images with artifacts. The results showed an improvement in the performance of the central model when opting to keep the images with artifact in the training databases.