Attribution Energy As a Companion Interpretability Metric for Grad-CAM In CBCT Image-Quality Assessment
Abstract
Purpose
Deep learning models have shown promise for image-quality rating tasks, yet the visual cues they rely on are often difficult to interpret. Grad-CAM and related saliency methods are commonly used to visualize attribution, but these heatmaps often yield unclear patterns and visual inspection alone can be misleading. To support repeatable, cross-case interpretation, we evaluate whether Attribution Energy (AE), a quantitative summary of Grad-CAM attribution, tracks expert-aggregated CBCT image quality ratings and can serve as a complementary interpretability signal.
Methods
A convolutional neural network (CNN)-based image quality prediction model was trained to predict consensus Likert-scale image quality scores (1–5) from axial head-and-neck CBCT slices across five categories: parotid, submandibular glands, larynx, mandible, and overall. Consensus scores were computed as a weighted average across expert raters. Grad-CAM maps were computed per slice, and AE was defined as the volumetric sum of raw Grad-CAM values across each CBCT volume. AE was computed for the CNN test set (16 CBCT volumes per category) and compared with consensus scores using volume-level Spearman rank correlation. Correlations were summarized across leave-one-out evaluations to assess sensitivity to individual volumes using median Spearman (ρ) and p-values.
Results
AE was significantly correlated with consensus score for parotid (median ρ = 0.71), submandibular gland (median ρ = 0.89), mandible (median ρ = 0.70), and overall (median ρ = 0.67) (all p ≤ 0.01). The larynx showed weaker, non-significant association (median ρ = 0.25, p = 0.37), suggesting structure-dependent behavior and divergence between model attribution and expert scoring cues.
Conclusion
This proof-of-concept study demonstrates that AE can complement Grad-CAM inspection by quantifying overall attribution strength, enabling cross-case comparison. As a companion metric, AE could be used to identify cases with atypical attribution behavior and support interpretability and quality assurance of deep learning-based CBCT image-quality assessment models.