<p>Interpretability in deep learning, particularly in convolutional neural networks (CNNs), remains a critical challenge due to the complexity of models obscuring decision-making processes. This study performs a rigorous comparative analysis of two key filter-level explanation strategies: removal-based and zeroing-based methods. The evaluation framework is grounded in fidelity, sparsity, robustness, and human interpretability. Extensive experiments were conducted using CNN architectures like ResNet-50 and VGG-16, across diverse medical imaging datasets, including brain tumors, prostatic hyperplasia, and breast cancer. The analysis reveals distinct behavioral patterns and practical implications for both strategies. Filter removal can sometimes mislead interpretations by overemphasizing or suppressing features, potentially altering model behavior. Conversely, zeroing-based methods better preserve structural fidelity, providing more stable and interpretable pathways for class-specific predictions. By leveraging activation visualizations and zeroing strategies, the study offers clearer insights into class predictions and decision processes. The work contributes a foundational understanding for integrating filter-based interpretability into practical AI development, especially for model debugging and deployment in clinical contexts. Overall, the findings advance understanding of CNN behavior and support the creation of trustworthy, human-aligned explanations essential in high-stakes decision-making environments such as healthcare.</p>