Deep learning has revolutionized artificial intelligence by enabling highly accurate models across diverse domains such as computer vision and autonomous driving. However, the inherent complexity of deep neural networks often results in black-box models, whose decision-making processes lack transparency and interpretability. Explainable Artificial Intelligence addresses this challenge by developing techniques that make model predictions more understandable to humans. Among these, Class Activation Mapping (CAM) and its variants: Grad-CAM, Grad-CAM++ and Score-CAM, have become pivotal in providing visual explanations for convolutional neural networks. This paper presents a systematic comparison of several prominent CAM-based explanation methods applied to the AlexNet architecture on a standardized Dogs vs. Cats image dataset. We evaluate these methods using quantitative metrics including Intersection over Union, Dice coefficient and Deletion and Insertion tests, aiming to reveal their respective strengths, limitations and practical utility. These findings offer valuable insights to guide practitioners in selecting appropriate explanation techniques. Specifically, we show that combining CAM-based techniques enables a deeper understanding of pixel importance by identifying critical regions composed of both object and background pixels and by guiding data preprocessing promotes more stable training and improves the model’s focus. Furthermore, we demonstrate the potential for constructing robust feature attribution strategies grounded in layer-wise analysis and assess the reliability of various XAI algorithms by evaluating their behavior across multiple threshold levels.