This study addresses key challenges in image recognition using Artificial Intelligence neural networks, specifically the lack of trust and limited interpretability caused by their “black-box” decision-making mechanisms. Unlike existing research, which often focuses on a single interpretability method or model performance optimization, this study systematically investigates visualization and trust-building strategies for high-risk applications. An enhanced Convolutional Neural Network (CNN) is developed by integrating Grad-CAM and LIME methods to explore interpretability techniques in visual image recognition. Evaluated on the CIFAR-10 dataset and a subset of ImageNet, the proposed model achieves recognition accuracies of 89.8% and 96.33%, respectively, marking a 3.2% improvement over the baseline AlexNet model. A novel dual-modal interpretability framework is introduced. Grad-CAM generates class-discriminative heatmaps to visualize global attention regions, while LIME provides localized feature attribution for individual predictions. The combination of these techniques significantly enhances model transparency, achieving an interpretation consistency score of 0.82. Experimental results show that the model maintains strong classification performance, with an F1 score of 92.2%. Insertion and deletion tests further confirm the reliability of the heatmap-based visualizations. To comprehensively evaluate interpretability, a multi-dimensional assessment system is implemented, combining quantitative metrics with expert human evaluation. This study provides valuable insights for the development of trustworthy AI systems, especially in high-stakes domains such as medical diagnostics and autonomous driving.
Zhang et al. (Thu,) studied this question.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: