A systematic review of active learning approaches in the selection of medical images

Background: Active Learning has been proven to be an effective way to maximize the model’s learning capacity, using fewer amounts of labeled data. In the field of medical imaging data, data and annotations can be scarce and very expensive to obtain, so techniques like Active Learning can be a useful solution. Methods: For this systematic review, the data sources were obtained through IEEE Explore, PubMed, and ACM Digital Library, between the period of 2018 and 2023. Only studies that belonged to the field of healthcare (using medical images as a dataset) and machine learning, written in English and that were not a book, or a survey were used. Covidence was used as a tool to synthesize the results. Results: From 336 records, 51 were included in this review. Interpretation: Most studies showed that Active Learning can have a positive impact on the construction of models, however, it is important to not consider only the informativeness/uncertainty of the sample, but also the distribution of the data, reducing the probability of selecting samples that are not representative enough of the dataset or outliers. Active Learning is usually an iterative process until a stop criterion is met, for example, the model’s performance. To evaluate an Active Learning solution, the proposed method is usually compared with random sampling, or other Active Learning queries.