Federated learning is an emerging technology that enables the decentralised training of machine learning-based methods for medical image analysis across multiple sites while ensuring privacy. This review paper thoroughly examines federated learning research applied to medical image analysis, outlining technical contributions. We followed the guidelines of Okali and Schabram, a review methodology, to produce a comprehensive summary and discussion of the literature in information systems. Searches were conducted at leading indexing platforms: PubMed, IEEE Xplore, Scopus, ACM, and Web of Science. We found a total of 433 papers and selected 118 of them for further examination. The findings highlighted research on applying federated learning to neural network methods in cardiology, dermatology, gastroenterology, neurology, oncology, respiratory medicine, and urology. The main challenges reported were the ability of machine learning models to adapt effectively to real-world datasets and privacy preservation. We outlined two strategies to address these challenges: non-independent and identically distributed data and privacy-enhancing methods. This review paper offers a reference overview for those already working in the field and an introduction to those new to the topic.

Advances in artificial intelligence (AI) have triggered interest in using intelligent systems to improve prenatal detection of fetal congenital heart defects (CHDs). Our aim is to systematically examine the current literature on diagnostic performance of AI-enabled prenatal cardiac ultrasound. This systematic review and meta-analysis was registered with PROSPERO (CRD42024549601). Embase, Medline, Cochrane Central Database of Controlled Trials, and CINAHL were searched from inception until February 2025. Studies evaluating AI performance in prenatal detection of fetal CHDs were eligible for inclusion, and studies focusing on the application of AI before 16 weeks of gestation, or using three- or four-dimensional ultrasound, were excluded. Pooled sensitivity and specificity were obtained using random-effect method, and pooled proportions using the Freeman-Tukey arcsine square root transformation. Heterogeneity was assessed with I² statistics. Risk of bias and adherence to reporting standards were assessed using QUADAS-2 and TRIPOD+AI, respectively. Risk of publication bias was assessed with Deek's test and certainty of evidence for outcomes with GRADE approach. Fifteen studies were included, of which fourteen developed and evaluated a model and one externally evaluated a previously trained model. Images and videos obtained during cardiac screening or fetal echocardiography of 30.121 fetuses were used for training, validation and testing. For the binary task of classifying heart as normal or abnormal, AI models achieved a pooled sensitivity of 0.89 (95% CI 0.83–0.93, I² = 77.92%) and specificity of 0.91 (95% CI 0.84–0.95, I² = 77.92%). The subgroup analysis showed that models tested on various CHDs exhibited lower sensitivity compared to those tested for a specific cardiac abnormality (0.85; 95% CI 0.75–0.91 vs 0.92; 95% CI 0.87–0.96), while specificity remained comparable (0.90; 95% CI 0.79–0.96 vs 0.91; 95% CI 0.81–0.97). Overall, AI models performed better than operators with lower expertise and were nearly comparable to experts; however, the human comparator group (median six clinicians, IQR 3–10) was usually small and non-blinded. Relevant sources of heterogeneity were the types of cardiac views collected, the prevalence of CHDs across different datasets, and the types of CHDs examined. The risk of bias was moderate-high and adherence to reporting standards low (>70% in 18/51 TRIPOD+AI items). The risk of publication bias was not statistically significant (Deek's test p = 0.474). These findings suggest that AI models perform better than clinicians with lower expertise, but this must be interpreted with caution due to the high risk of bias and sources of heterogeneity. This study was partly supported by the InnoHK-funded Hong Kong Centre for Cerebro-cardiovascular Health Engineering (COCHE) Project 2.1 (Cardiovascular risks in early life and fetal echocardiography). ATP and JAN are supported by the 10.13039/501100000272 National Institute for Health and Care Research (NIHR) 10.13039/501100013373 Oxford Biomedical Research Centre (BRC).

Activity recognition, a key component in pervasive healthcare monitoring, relies on classification algorithms that require labeled data of individuals performing the activity of interest to train accurate models. Labeling data can be performed in a lab setting where an individual enacts the activity under controlled conditions. The ubiquity of mobile and wearable sensors allows the collection of large datasets from individuals performing activities in naturalistic conditions. Gathering accurate data labels for activity recognition is typically an expensive and time-consuming process. In this paper we present two novel approaches for semi-automated online data labeling performed by the individual executing the activity of interest. The approaches have been designed to address two of the limitations of self-annotation: (i) The burden on the user performing and annotating the activity, and (ii) the lack of accuracy due to the user labeling the data minutes or hours after the completion of an activity. The first approach is based on the recognition of subtle finger gestures performed in response to a data-labeling query. The second approach focuses on labeling activities that have an auditory manifestation and uses a classifier to have an initial estimation of the activity, and a conversational agent to ask the participant for clarification or for additional data. Both approaches are described, evaluated in controlled experiments to assess their feasibility and their advantages and limitations are discussed. Results show that while both studies have limitations, they achieve 80% to 90% precision.

Monitoring the healthy development of a fetus requires accurate and timely identification of different maternal-fetal structures as they grow. To facilitate this objective in an automated fashion, we propose a deep-learning-based image classification architecture called the COMFormer to classify maternal-fetal and brain anatomical structures present in two-dimensional fetal ultrasound images. The proposed architecture classifies the two subcategories separately: maternal-fetal (abdomen, brain, femur, thorax, mother's cervix, and others) and brain anatomical structures (trans-thalamic, trans-cerebellum, trans-ventricular, and non-brain). Our proposed architecture relies on a transformer-based approach that leverages spatial and global features by using a newly designed residual cross-variance attention (R-XCA) block. This block introduces an advanced cross-covariance attention mechanism to capture a long-range representation from the input using spatial (e.g., shape, texture, intensity) and global features. To build COMFormer, we used a large publicly available dataset (BCNatal) consisting of 12, 400 images from 1,792 subjects. Experimental results prove that COMFormer outperforms the recent CNN and transformer-based models by achieving 95.64% and 96.33% classification accuracy on maternal-fetal and brain anatomy, respectively.