Deep learning (DL) has become an essential tool in medical imaging. However, privacy regulations often make it difficult to collect and share medical data, limiting the availability of large and diverse datasets needed to train robust models. Federated learning (FL) offers a solution by enabling collaborative model training across multiple institutions without compromising patient privacy. However, the impact of the number of clients and of unbalanced data distribution on model performance has received limited attention.This study investigates the performance of FL compared to centralized learning (CL) in the tasks of medical image classification and segmentation under different client setups. For the classification task, we used chest X-ray (CXR) images to classify them into one of four categories using ResNet18. For the segmentation task, we used cardiac computed tomography (CT) images to train a 3D U-Net model to segment the left ventricle (LV) and the myocardium (Myo). For both tasks, we evaluated the efficacy of FL with 2, 4 and 8 clients with and without data unbalance, and an additional 16 clients setup for segmentation. In classification, the trained model achieved an accuracy of 0.93 in the centralized approach, with similar performance in federated settings.Performance remained consistent across balanced and unbalanced data distributions, with sensitivity consistently around 0.94. For segmentation, the centralized model yielded a Dice Similarity Coefficient (DSC) of 0.877 (Myo) and 0.926 (LV), and the FL approach models yielded a DSC of 0.867 (Myo) and 0.922 (LV); 0.867 (Myo) and 0.923 (LV); 0.863 (Myo) and 0.921 (LV) for 2, 4 and 8 clients respectively.Our results show that FL achieves competitive performance compared to CL in medical image classification and segmentation tasks. Moreover, the experiments involving unbalanced data did not reveal any significant decrease in performance. For the segmentation task, the federated approach achieved consistent results for both structures across different client setups, closely matching the performance of the centralized model. These results suggest that FL is a viable alternative to CL, especially in scenarios where privacy and decentralization are important.