FXG Score-CAM: Comprehensive Score-Weighted Visual Explanations for Convolutional Neural Networks

Convolutional Neural Networks (CNNs) have shown impressive performance on solving vision problems, but their limited interpretability is a concern, especially in high-stakes applications where trust and clarity are essential. Among post-hoc visual explanation techniques, Score-CAM is widely used, but has notable limitations. It relies solely on activation maps, ignoring input and per-neuron sensitivity. Additionally, this method typically extracts information from a single layer (usually the final convolutional layer) making its output highly dependent on layer selection while disregarding valuable contributions from other layers. Score-CAM also applies uniform spatial weighting within feature maps, failing to capture variations in importance across individual spatial locations. To address these shortcomings, we propose FXG (FullXGrad) Score-CAM, a novel interpretability method that integrates FullGrad, XGrad-CAM, and Score-CAM. FXG Score-CAM integrates activation, input, and bias attribution into the saliency map using the Score-CAM weighting method, aggregating information across all layers via the FullGrad framework while leveraging XGrad-CAM for spatially adaptive feature weighting. Our experimental evaluations on benchmark datasets demonstrate that FXG Score-CAM consistently outperforms Score-CAM and other state-of-the-art explainability methods in both qualitative and quantitative assessments. It also passes standard sanity checks, confirming explanation reliability. Code is available at https://github.com/quannguyen2153/fxg-score-cam .