Cardiac magnetic resonance imaging (MRI) segmentation is an essential aspect of quantitative cardiovascular analysis, facilitating accurate evaluation of ventricular volumes, myocardial mass, and functional parameters. Deep learning-based segmentation models have shown strong performance on benchmark datasets such as ACDC, but they remain challenging to deploy in real-world multi-centre settings. Data privacy laws make it hard to share data across institutions, and differences in imaging protocols and patient populations mean that data is not always distributed in the same way (non-IID). This can have a big impact on how well models work together and how well they generalise. To address these issues, we first evaluate advanced segmentation architectures, including UNet++ and FPN with EfficientNet-based encoders, and assess multiple hybrid combinations at the probability level. We further improve the ensemble strategy by using a genetic algorithm to automatically identify the optimal model-weighting scheme, rather than fixed combination coefficients. The genetic algorithm explores the solution space to identify the optimal weight configuration based on segmentation metrics. The best hybrid configuration is then chosen as the input architecture for the federated learning stage. We propose a privacy-preserving federated ensemble framework that enables multiple clients to collaboratively train segmentation models without sharing raw MRI data. We methodically evaluate three federated optimisation strategies: FedAvg under IID and non-IID client distributions, and FedProx, which incorporates proximal regularisation to reduce client drift. The genetically optimised ensemble is always used in all federated setups. A thorough analysis of ACDC testing volumes employing overlap- and boundary-based metrics illustrates that the amalgamation of hybrid learning with genetic optimisation and federated training enhances robustness in heterogeneous environments while maintaining data confidentiality, thus providing an efficient approach for secure multi-centre cardiac MRI segmentation.

Breast cancer is still one of the leading causes of death in women, and finding it early is essential for treatment to work. This study presents a sizeable deep learning architecture for classifying and segmenting breast ultrasound images. It utilizes federated learning to maintain patients' data privacy. The proposed pipeline begins with a thorough preprocessing stage that includes scaling, normalizing, and advanced image augmentation using affine and contrast-based changes. We employ four convolutional neural network architectures for hierarchical classification: ResNet50, EfficientNet, VGG16, and Xception. First, we distinguish between typical cases and abnormal ones. We then further classify abnormal images into benign and malignant classes. We employ an ensemble technique that combines the outputs of ResNet50 and EfficientNet through a weighted average optimized by a genetic algorithm to enhance the model's resilience. This method dramatically improves the classification's effectiveness, achieving higher accuracy and reliability. We use a federated learning system with the federated averaging (FedAvg) algorithm to improve data privacy. Our federated architecture maintains high accuracy while ensuring that the raw data stays at its local source. We test it with both single-client and multi-client setups. Ultimately, we employ a hybrid architecture that combines the feature maps of ResNet50 and EfficientNet to segment images of lesions known to be malignant. This yields significant spatial agreement with expert annotations. The Dice score and intersection over union (IoU) are two evaluation criteria that demonstrate the effectiveness of our segmentation model. This all-in-one system offers accurate and privacy-conscious breast ultrasound analysis, indicating that it could be beneficial in decentralized healthcare settings. The suggested method was tested using a standard breast magnetic resonance imaging (MRI) dataset, demonstrating its robustness and applicability in various situations. We used the accuracy, precision, recall, and F1-score to measure performance, and we found that the classification was accurate up to 96%. This paper discusses a scalable system that maintains people's privacy by utilizing ensemble learning, optimization-driven feature selection, and federated learning to aid doctors in early breast cancer detection using MRI data. This will make it easier for doctors to use this system in a broader range of medical tests.