Quantifying Explanation Consistency: The C-Score Metric for CAM-Based Explainability in Medical Image Classification

Kabilan Elangovan; Daniel Ting

cs.CVApr 9, 2026

Quantifying Explanation Consistency: The C-Score Metric for CAM-Based Explainability in Medical Image Classification

Authors: Kabilan Elangovan, Daniel Ting

Organizations: Singapore Health Services, Singapore · Singapore Eye Research Institute, Singapore

Abstract

Class Activation Mapping (CAM) methods are widely used to generate visual explanations for deep learning classifiers in medical imaging. However, existing evaluation frameworks assess whether explanations are correct, measured by localisation fidelity against radiologist annotations, rather than whether they are consistent: whether the model applies the same spatial reasoning strategy across different patients with the same pathology. We propose the C-Score (Consistency Score), a confidence-weighted, annotation-free metric that quantifies intra-class explanation reproducibility via intensity-emphasised pairwise soft IoU across correctly classified instances. We evaluate six CAM techniques: GradCAM, GradCAM++, LayerCAM, EigenCAM, ScoreCAM, and MS GradCAM++ across three CNN architectures (DenseNet201, InceptionV3, ResNet50V2) over thirty training epochs on the Kermany chest X-ray dataset, covering transfer learning and fine-tuning phases. We identify three distinct mechanisms of AUC-consistency dissociation, invisible to standard classification metrics: threshold-mediated gold list collapse, technique-specific attribution collapse at peak AUC, and class-level consistency masking in global aggregation. C-Score provides an early warning signal of impending model instability. ScoreCAM deterioration on ResNet50V2 is detectable one full checkpoint before catastrophic AUC collapse and yields architecture-specific clinical deployment recommendations grounded in explanation quality rather than predictive ranking alone.

Figures & tables

Symbol	Definition
$C(c,\theta,m)$	C-Score for class $c$ , checkpoint $\theta$ , method $m$ ; $C\in[0,1]$
$G(c,\theta)$	Gold list: correctly classified test images for class $c$ at $\theta$ under $\tau{=}0.5$
$H_{i}^{(c,\theta,m)}$	Normalised CAM heatmap; $H_{i}\in[0,1]^{W\times H}$
$p_{i}^{(c,\theta)}$	Sigmoid confidence for class $c$ on image $i$
$w_{i}$	Confidence weight: $w_{i}=p_{i}/\sum_{j}p_{j}$ (normalised over gold list)
$s_{\text{IoU}}(H_{i},H_{j})$	Soft-IoU: $\sum\min(H_{i},H_{j})/\sum\max(H_{i},H_{j})$ , element-wise

Table 1: C-Score notation and definitions.

Appendix figures & tables11 assets

Supplementary material from the paper’s appendix.

Appendix

		DenseNet201		InceptionV3		ResNet50V2
Ep.	Phase	AUC	Acc.	AUC	Acc.	AUC	Acc.
1	TL	0.9184	61.60%	0.8705	66.13%	0.9582	81.48%
2	TL	0.9656	84.64%	0.9310	64.33%	0.9760	94.20%
3	TL	0.9657	45.48%	0.9509	69.97%	0.9807	82.42%
4	TL	0.9804	90.96%	0.9542	86.35%	0.9800	94.71%
5	TL	0.9848	94.03%	0.9592	78.33%	0.9876	95.82%

Appendix

Table 2: Complete 30-epoch test AUC and accuracy for all three architectures. TL = Transfer Learning (epochs 1–20); FT = Fine-Tuning (epochs 21–30). Red : AUC

<0.50

(mode collapse). Amber : accuracy

<50\%

with intact AUC (threshold-crossing, rankings preserved). DenseNet201 epochs 21–23 exhibit boundary-reorganisation accuracy collapse while AUC remains high.

Technique	E1	E5	E10	E15	E20 (TL End)	E25 (Mid-FT)	E30 (FT End)
GradCAM	0.113	0.168	0.170	0.198	0.197	0.744	0.807
GradCAM++	0.358	0.461	0.546	0.566	0.549	0.916	0.870
LayerCAM	0.403	0.479	0.569	0.583	0.559	0.915	0.871
ScoreCAM	0.460	0.563	0.622	0.632	0.630	0.933	0.880
EigenCAM	0.635	0.634	0.635	0.636	0.635	0.908	0.846
MS-GradCAM++	0.322	0.380	0.439	0.460	0.445	0.644	0.618

Appendix

Table 3: Global weighted C-Score for DenseNet201. Green : C-Score

\geq 0.80

; amber :

<0.30

; red :

0.000

. Step-change improvement at TL

\to

FT transition observable across all methods.

Technique	E1	E5	E10	E15	E20 (TL End)	E25 (Mid-FT)	E30 (FT End)
GradCAM	0.169	0.242	0.209	0.195	0.196	0.875	0.244
GradCAM++	0.475	0.508	0.492	0.484	0.485	0.808	0.762
LayerCAM	0.567	0.583	0.572	0.567	0.568	0.802	0.763
ScoreCAM	0.392	0.386	0.383	0.381	0.379	0.790	0.759
EigenCAM	0.758	0.759	0.758	0.757	0.756	0.896	0.852
MS-GradCAM++	0.419	0.417	0.417	0.415	0.419	0.659	0.654

Appendix

Table 4: Global weighted C-Score for InceptionV3.

Technique	E1	E5	E10	E15	E20 (TL End)	E25 (Mid-FT)	E30 (FT End)
GradCAM	0.422	0.387	0.573	0.400	0.385	0.272	0.370
GradCAM++	0.320	0.613	0.642	0.607	0.593	0.676	0.478
LayerCAM	0.513	0.667	0.681	0.662	0.654	0.675	0.489
ScoreCAM	0.517	0.698	0.629	0.621	0.612	0.014	0.000
EigenCAM	0.589	0.685	0.693	0.692	0.688	0.678	0.495
MS-GradCAM++	0.313	0.508	0.527	0.508	0.507	0.533	0.409

Appendix

Table 5: Global weighted C-Score for ResNet50V2. Net E20

\to

E30 change negative for majority of techniques. ScoreCAM trajectory

0.612\to 0.014\to 0.000

provides pre-collapse early-warning signal one checkpoint before AUC collapse at E30.

Figure 1: Global weighted C-Score trajectory across training phases.

Figure 2: C-Score heatmap comparison at transfer learning end (E20) and fine-tuning end (E30).

Figure 3: Net C-Score change (E30 $-$ E20) by architecture and method.

Technique	E1	E5	E10	E15	E20 (TL)	E25 (FT)	E30 (FT)
Normal (Class 0)
GradCAM	0.159	0.593	0.606	0.663	0.664	0.000	0.924
GradCAM++	0.424	0.669	0.694	0.728	0.718	0.000	0.918
LayerCAM	0.471	0.672	0.696	0.729	0.716	0.000	0.919
ScoreCAM	0.539	0.621	0.688	0.738	0.714	0.000	0.907
EigenCAM	0.680	0.682	0.688	0.689	0.688	0.000	0.895

Appendix

Table 6: Per-class C-Score for DenseNet201. E25 shows 0.000 for all Normal-class techniques (gold-list population effect, not method failure): no Normal images pass

\tau=0.5

at this checkpoint during boundary reorganisation.

Technique	E1	E5	E10	E15	E20 (TL)	E25 (FT)	E30 (FT)
Normal (Class 0)
GradCAM	0.047	0.288	0.380	0.317	0.335	0.840	0.847
GradCAM++	0.469	0.542	0.536	0.516	0.509	0.872	0.851
LayerCAM	0.613	0.642	0.631	0.620	0.623	0.866	0.852
ScoreCAM	0.292	0.332	0.303	0.286	0.286	0.845	0.852
EigenCAM	0.770	0.777	0.775	0.774	0.774	0.938	0.922

Appendix

Table 7: Per-class C-Score for InceptionV3. GradCAM Pneumonia achieves 0.887 at E25 but collapses to 0.008 at E30, reflecting non-monotonic attribution dynamics post fine-tuning.

Technique	E1	E5	E10	E15	E20 (TL)	E25 (FT)	E30 (FT)
Normal (Class 0)
GradCAM	0.333	0.095	0.573	0.455	0.481	0.798	0.000
GradCAM++	0.364	0.656	0.662	0.643	0.634	0.798	0.000
LayerCAM	0.517	0.715	0.711	0.696	0.688	0.798	0.000
ScoreCAM	0.508	0.882	0.652	0.645	0.649	0.041	0.000
EigenCAM	0.593	0.745	0.727	0.728	0.723	0.802	0.000

Appendix

Table 8: Per-class C-Score for ResNet50V2. Normal class registers 0.000 across all techniques at E30 (mode collapse). Pneumonia GradCAM at E25 drops to 0.000 while GradCAM++ and LayerCAM remain above 0.600.

Figure 4: Per-class C-Score trajectory by architecture.

Explore similar work

Class Activation Mapping in Explainable Computer Vision: A Method-Centered Review of CNN, Transformer, and Foundation-Model-Era Visual Explanations
Aug 12, 2026AmirHossein Eshghi, Hamid Saadatfar, Seyyed Ali Hoseini +2Transformer Interpretability Feature Attribution
Department of Computer Engineering, University of Birjand, Iran
Med-CAM: Minimal Evidence for Explaining Medical Decision Making
Apr 15, 2026Pirzada Suhail, Aditya Anand, Amit SethiExplainable Image Classification Explainable Medical Image Analysis
IIT Bombay
How to Evaluate and Refine your CAM
May 14, 2026Luca Domeniconi, Alessandra Stramiglio, Michele Lombardi +1Feature Attribution Explainability Evaluation
University of Bologna, Bologna, Italy

Aug 12, 2026J/K move · Enter open · S save

Class Activation Mapping in Explainable Computer Vision: A Method-Centered Review of CNN, Transformer, and Foundation-Model-Era Visual Explanations

AmirHossein Eshghi, Hamid Saadatfar, Seyyed Ali Hoseini, AmirMohsen Eshghi+1

Department of Computer Engineering, University of Birjand, Iran

Transformer Interpretability Feature Attribution Explainable Image Classification

Open paper AskPDF arXiv

Class activation mapping (CAM) is one of the most widely used visual explanation families in explainable artificial intelligence. Its purpose is intuitive: it converts internal model evidence into a heatmap that highlights the image regions, convolutional channels, tokens, or patches that support a target class or concept. Since the first CAM formulation in 2016, the field has moved far beyond global-average-pooled CNN classifiers. CAM-style methods now include gradient-based post-hoc explanations, gradient-free score and ablation methods, high-resolution upscaling, weakly supervised localization and segmentation, transformer token attribution, causal and debiasing methods, and foundation-model-era approaches that use CLIP, DINO, SAM, or feature-distribution comparisons. This review synthesizes a strict corpus of 57 method-centered papers published from 2016 onward. The paper develops a taxonomy that separates methods by attribution mechanism, architectural dependence, and evaluation objective. It then reviews gradient-based CAMs, recent and hybrid CAM-style methods, and model-based or architecture-aware methods. Across the corpus, the main trend is clear: the field is shifting from explaining one class score in one low-resolution CNN layer toward comparative, multi-layer, probabilistic, token-aware, and foundation-model-aware explanations. At the same time, evaluation remains fragmented. Faithfulness, localization, robustness, computational cost, and human trust are often measured with different protocols. The review therefore emphasizes not only what each method contributes, but also which gap it leaves open and which later methods attempt to close that gap.

Abstract

Explore similar work

Class Activation Mapping in Explainable Computer Vision: A Method-Centered Review of CNN, Transformer, and Foundation-Model-Era Visual Explanations

Med-CAM: Minimal Evidence for Explaining Medical Decision Making

How to Evaluate and Refine your CAM