Modern imaging technologies generate thousands of potential features that can overwhelm computational resources and lead to unreliable models through overfitting. In this context, feature selection methods are essential for identifying the most relevant imaging characteristics while discarding redundant or irrelevant data. This review examines the most adequate methods according to the state-of-the-art. Then, we perform a comparative analysis that evaluates these methods across key dimensions: computational efficiency, ability to capture feature interactions, model dependency, and suitability for high-dimensional or multi-modal datasets. Rather than cataloging technical details, we emphasize each approach’s fundamental principles and practical trade-offs. Our synthesis provides researchers and clinicians with a practical framework for selecting appropriate feature selection strategies that balance computational efficiency, model performance, and clinical interpretability–ultimately supporting the development of more robust diagnostic tools that can meaningfully improve patient care.