Quantifying Explanation Consistency: The C-Score Metric for CAM-Based Explainability in Medical Image Classification
Kabilan Elangovan 1 2 Daniel Ting 1 2 1
Singapore Health Services, Singapore 2 Singapore Eye Research Institute, Singapore
arXiv:2604.08502v1 [cs.CV] 9 Apr 2026
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.
1
Introduction
1.1
The Clinical Deployment Gap in Medical AI
on “subvisual” or “nonvisual” statistical patterns—features mathematically present but imperceptible to radiologists— creating a fundamental verification gap where physicians cannot visually confirm the model’s reasoning [11].
Deep learning models for medical image analysis have achieved, and in some domains exceeded, expert-level discriminative performance [1, 2]. Convolutional neural networks trained on chest X-ray datasets now achieve AUC values exceeding 0.99 for pneumonia detection [3], and similar performance has been reported across ophthalmology, dermatology, and radiology [4].
1.2
The Reliability Problem: CAM Methods Fail Basic Sanity Checks
Class Activation Mapping methods were introduced to address the deployment trust gap by making CNN spatial reasoning visible and auditable [12]. However, a series of foundational studies has revealed that CAM methods themselves suffer from reliability failures that are independent of — and invisible to — classification performance metrics.
However, a fundamental tension exists between classification performance and clinical trustworthiness. High AUC certifies that a model correctly ranks pathological cases above normal ones in terms of predicted probability. It does not certify what image features the model is using to make that ranking, nor whether those features correspond to genuine pathological findings. The “Clever Hans” phenomenon — where models exploit dataset-specific shortcuts rather than genuine pathological features — is now well-documented across medical imaging domains [5, 6, 7]. Chest X-ray classifiers have been shown to leverage equipment markers, patient positioning artefacts, and pacemaker presence as proxy discriminators [8, 9, 10]. Moreover, recent perspectives highlight that AI often relies
Model Parameter Randomization [13]: Several widely used explanation methods produce nearly identical heatmaps for a fully trained model and for a model with randomly re-initialised weights. If an explanation method cannot distinguish between a trained model and a random one, it is not measuring the model’s learned reasoning. Input Invariance Problem [14]: A constant shift applied to all input pixels — a transformation with exactly zero effect on model predictions — produces completely different 1
Quantifying Explanation Consistency: The C-Score Metric for CAM-Based Explainability in Medical Image Classification
saliency attributions in many popular methods.
liable mental model of a system whose attention is arbitrarily variable. As Ghorbani et al. [15] noted, even if a pathology predictor is robust, a fragile interpretation would still be highly disconcerting if a clinician is using that interpretation to guide clinical decision-making. Furthermore, as McLeod et al. [11] note, because deep learning models lack causal common sense and may diverge from human visual search patterns, ensuring that these divergent strategies are at least consistent across patient populations is a necessary safeguard against spurious correlations. Yet, the field currently lacks a standardized, quantifiable metric to evaluate this fundamental requirement.
Interpretation Fragility [15]: Imperceptible adversarial perturbations bounded by L∞ =2 — leaving model predictions unchanged — can redirect spatial attention to arbitrary image regions. Faithfulness vs. Visual Appeal [16]: GradCAM’s spatial pooling of gradients causes it to highlight regions larger than those the model actually uses, creating visually appealing but spatially imprecise attributions. 1.3
Gaps in Current CAM Evaluation Approaches
The recognition of CAM reliability failures has stimulated research into quantitative evaluation frameworks, yet four systematic gaps remain.
1.5
We propose the C-Score (Consistency Score) as a metric that directly addresses the intra-class consistency gap. The core thesis is that explainability itself cannot be proven because there is no objective ground truth for what a “correct” explanation looks like, but consistency in explainability can be quantified. A model that consistently attends to the same anatomical regions across patients with the same diagnosis demonstrates at minimum that its explanations are reproducible. This reproducibility is a necessary precondition for clinical interpretability.
First, perturbation-based metrics (deletion and insertion curves) create out-of-distribution masked inputs and conflate attribution quality with model behaviour on corrupted data [17, 18]. Second, localisation metrics compare CAM-predicted attention regions to ground-truth annotations. The pointing game and IoU-with-bounding-box require expensive perimage annotation unavailable at scale and are not applicable across training checkpoints.
The C-Score is formulated as a confidence-weighted, intensity-emphasised mean pairwise soft IoU across correctly classified instances of the same class at a given model checkpoint. The metric is designed specifically for image classification tasks trained with image-level labels and no pixel-level annotations, where spatial explanations are derived post hoc using CAM-based attribution methods. Under this setting, C-Score provides an annotation-free measure of spatial explanation consistency that can be computed at every training epoch and applied across different CAM techniques and CNN architectures.
Third, and most critically: no existing evaluation framework addresses intra-class consistency. Current metrics ask whether the model attends to the correct spatial region on a per-image basis. The fundamentally different clinical question — does the model consistently apply the same visual reasoning strategy across different patients with the same diagnosis? — remains unanswered. Fourth, the disagreement problem extends to the choice of evaluation metric itself: different faithfulness metrics select different methods as “best” on the same model and data [19, 20]. 1.4
The C-Score: Contribution and Positioning
This paper reports five primary contributions: (i) Formal specification of the C-Score metric with complete mathematical definition, intensity emphasis rationale (α=2.0), confidence weighting, and gold-list formation under threshold τ =0.5.
Why Consistency is the Fundamental Clinical Requirement
The traditional assumption has been that explainability is equivalent to highlighting the clinically correct region of interest. A critical reconceptualisation follows from the reliability failures documented above. Clinical ROI alignment is desirable, but is not the fundamental requirement for trust. What is fundamental is whether the model consistently applies its learned decision strategy across similar cases.
(ii) Comprehensive evaluation across six CAM techniques, three CNN architectures, and thirty training epochs on the Kermany chest X-ray dataset, revealing both intra-phase and inter-phase consistency dynamics. (iii) Identification and characterisation of three distinct mechanisms of AUC–consistency dissociation invisible to standard classification metrics.
Consistent “wrong” focus is addressable. Inconsistent focus is unpredictable. In a clinical deployment context, a physician can learn to interpret and mentally correct for a systematic bias in AI attention; they cannot develop a re-
(iv) Empirical demonstration of C-Score as a precollapse monitoring signal, with ScoreCAM deterioration on ResNet50V2 detected one training check2
Quantifying Explanation Consistency: The C-Score Metric for CAM-Based Explainability in Medical Image Classification
point before catastrophic AUC collapse.
Lago et al. [35] FDA-aligned consistency dimension and differs from HAAS [30] in being annotation-free and continuously computable across the training trajectory.
(v) Architecture- and technique-specific clinical deployment recommendations grounded in both AUC and C-Score trajectory evidence.
2
Related Work
2.1
CAM Methods for Medical Image Explainability
3.1
Experimental Setup
Training followed a deliberate two-phase protocol. Phase 1 — Transfer Learning (epochs 1–20): frozen backbone weights, classification head trained with the Adam optimiser [40] at lr = 1×10−4 with cosine annealing. Phase 2 — Fine-Tuning (epochs 21–30): all layers unfrozen at lr = 1×10−5 with label smoothing (ε=0.1) and gradient clipping (norm = 1.0). 3.2
Reviews of CAM methods in medical imaging (2024– 2025) have consistently identified the lack of standardised evaluation as the primary obstacle to clinical adoption [25, 26]. Both van der Velden et al. [27] and Suara et al. [28] conclude that consistency, alongside faithfulness, must be a primary evaluation criterion.
3.2.1
The C-Score: A Consistency Metric for CAM-Based Explanations M OTIVATION
The dominant evaluation paradigm in medical AI — AUCROC — quantifies discriminative ranking ability but provides no information about how the model reaches its decisions. A model achieving AUC = 0.99 may do so through spurious correlations with acquisition artefacts or demographic proxies. CAM methods were introduced to expose spatial reasoning patterns [12]. However, existing CAM evaluations focus on localisation fidelity and do not assess whether explanations are reproducibly consistent across different patients with the same pathology. C-Score fills this gap: rather than asking where the model attends, it asks how consistently it attends to that location across the clinical population.
Evaluation of Explanation Quality
Perturbation-based metrics [17, 18], localisation metrics [29], and human-grounded assessments [30] constitute the dominant evaluation paradigms. The ERASER benchmark [31] and Quantus toolkit [32, 33] provide multi-metric evaluation but do not address intra-class consistency. The disagreement problem — different faithfulness metrics yield conflicting rankings of explanation methods — has been documented by Krishna et al. [19] and confirmed across multiple benchmarks [20]. 2.3
Methodology
Experiments were conducted on the Kermany chest X-ray dataset [3], comprising 5,856 images with a test split of 317 Normal and 855 Pneumonia images. Three CNN architectures were evaluated: DenseNet201 [36] (20M parameters), InceptionV3 [37] (24M parameters), and ResNet50V2 [38] (25M parameters), each initialised from ImageNet [39] pretrained weights.
Selvaraju et al. [12] introduced GradCAM as a classdiscriminative visualisation method anchored to the final convolutional layer’s gradient-weighted activation. The method produced substantially more useful explanations than prior pixel-wise gradient methods and achieved wide adoption in medical imaging contexts. Subsequent gradient-based methods refined spatial attribution precision: GradCAM++ [21] introduced second-order gradient weighting; LayerCAM [22] preserved full spatial resolution through pixel-wise gradient–activation products. Gradient-free alternatives such as ScoreCAM [23] and EigenCAM [24] avoid gradient pathologies at increased computational cost. Multi-scale aggregation strategies have been proposed to balance semantic strength at deep layers with spatial resolution at shallow layers [21].
2.2
3
3.2.2
F ORMAL D EFINITION
Let a CNN classifier f (x; θ) be parameterised by weights θ with sigmoid output p = σ(f (x; θ)) ∈ (0, 1). The gold list G(c, θ) for class c at checkpoint θ is:
Consistency and Reproducibility in Prior Literature
The closest existing work to C-Score is the Difference of Means (DoM) metric proposed by Ozer et al. [34], which measures consistency of saliency detectors across different network architectures by comparing mean activation maps. DoM addresses inter-architecture consistency. C-Score addresses intra-class consistency: whether the same model consistently attends to the same regions for different patients with the same diagnosis. C-Score complements the
(c,θ) G(c, θ) = i ∈ Dtest : yi = c ∧ pi ≥τ ,
τ = 0.5. (1)
Soft-IoU between two normalised heatmaps Hi , Hj ∈ [0, 1]W ×H : P u,v min Hi (u, v), Hj (u, v) . (2) sIoU (Hi , Hj ) = P u,v max Hi (u, v), Hj (u, v) 3
Quantifying Explanation Consistency: The C-Score Metric for CAM-Based Explainability in Medical Image Classification Table 1. C-Score notation and definitions. Symbol
lack consistency. At the target layer ℓ∗ , gradients therefore reflect a domain-incomplete pathway. Consequently, two pneumonia patients activating different ImageNet-derived features are likely to produce uncorrelated gradients, predicting a low initial C-Score for gradient-based attribution methods.
Definition
C(c, θ, m)
C-Score for class c, checkpoint θ, method m; C ∈ [0, 1] G(c, θ) Gold list: correctly classified test images for class c at θ under τ =0.5 (c,θ,m) Hi Normalised CAM heatmap; Hi ∈ [0, 1]W ×H (c,θ) pi Sigmoid confidence for class c on image i P wi Confidence weight: wi = pi / j pj (normalised over gold list) sIoU (Hi , Hj ) P Soft-IoU: P min(Hi , Hj )/ max(Hi , Hj ), element-wise Ĥi = Hiα Intensity emphasis; α = 2.0 Cglobal (θ, m) Support-weighted global C-Score across classes τ = 0.5 Classification threshold for gold list membership ℓ∗ Target layer: DenseNet201→conv5_block32_concat; InceptionV3→mixed10; ResNet50V2→conv5_block3_out
In contrast, fine-tuning is expected to propagate chest Xray–specific gradients through the entire backbone, aligning internal activations toward domain-relevant feature axes and thereby increasing cross-patient heatmap similarity. EigenCAM, which is entirely gradient-free, should theoretically remain unaffected by this phenomenon. 3.2.4
C-Score complements the Lago et al. [35] FDA-aligned consistency dimension and differs from HAAS [30] in being annotation-free and continuously computable. It operationalises a property not captured by deletion/insertion metrics, pointing game, or IoU with bounding box, each of which requires per-image evaluation rather than intra-class population assessment. 3.3
Intensity emphasis (α = 2.0 suppresses diffuse background): Ĥi = Hiα , α = 2.0. (3)
GradCAM [12]: Weighted linear combination of feature maps using globally average-pooled class-specific gradiP P ents: αkc = Z1 i,j ∂y c /∂Akij ; Lc = ReLU( k αkc · Ak ). Reference baseline method. Global pooling discards spatial gradient structure, producing coarse attribution blobs.
(c,θ)
pi
. (c,θ) j∈G(c,θ) pj
(4)
Per-class C-Score (confidence-weighted mean pairwise soft-IoU): X (wi + wj ) · sIoU (Ĥi , Ĥj ) C(c, θ, m) =
i,j∈G(c,θ) i<j
Z
,
GradCAM++ [21]: Spatially resolved second-order gradiP ent weighting: αkc = i,j wkij,c · ReLU(∂y c /∂Akij ). More effective for multi-instance patterns; consistently outperforms GradCAM in C-Score due to finer spatial discrimination.
(5)
where Z is the normalisation constant over all pairs.
LayerCAM [22]: Element-wise product of positive gradients and activations: Lc = ReLU(∂y c /∂A) ⊙ A. Retains full spatial resolution, producing compact attribution footprints. Achieves C-Score values comparable to GradCAM++ across all architectures.
Global support-weighted C-Score: Cglobal (θ, m) =
X
|G(c, θ)| · C(c, θ, m). ′ c′ |G(c , θ)|
P c
(6)
EigenCAM [24]: SVD of the activation tensor Aflat = U ΣV ⊤ ; first right singular vector v1 reshaped to spatial attribution. Entirely gradient-free: insensitive to gradient noise, saturation, and sparsity. Achieves uniquely stable CScore throughout transfer learning (DenseNet201: 0.635 at E1, 0.635 at E20; InceptionV3: 0.758 at E1, 0.756 at E20).
Table 1 summarises the notation. 3.2.3
CAM Methods Evaluated
Six CAM techniques were evaluated, applied to fixed target layers (DenseNet201: conv5_block32_concat; InceptionV3: mixed10; ResNet50V2: conv5_block3_out).
Confidence weighting: wi = P
R ELATIONSHIP TO E XISTING F RAMEWORKS
T HEORETICAL M OTIVATION : T HE G RADIENT F LOW–C ONSISTENCY C ONNECTION
We hypothesize that during transfer learning, frozen backbone weights restrict gradient signals to flow only through the classification head, resulting in early explanations that
ScoreCAM [23]: Per-channel input masking with forwardpass class probability measurement: wk = f c (X ⊙ M k ) − 4
Quantifying Explanation Consistency: The C-Score Metric for CAM-Based Explainability in Medical Image Classification
f c (Xbaseline ). Gradient-free; O(C) forward passes; subsampling at max_N=32 applied. Its collapse trajectory on ResNet50V2 provides the study’s clearest pre-collapse early-warning signal.
4.2
Tables 3–5 present the global weighted C-Score at seven checkpoints for all three architectures. Figure 1 visualises the trajectories; Figure 2 compares heatmaps at E20 vs. E30.
MS-GradCAM++ (Multi-Scale GradCAM++): To address the trade-off between semantic strength at deep layers and spatial resolution at shallow layers, we implement a multi-scale aggregation strategy. We compute GradCAM++ heatmaps at K distinct points in the feature hierarchy (e.g., for DenseNet201: conv3_block12, pool4, and conv5_block32) and compute their pixelwise arithmetic mean: LcMS =
1 X ↑ LcGradCAM++ (ℓ) , K
4.2.1
(7)
where ↑ (·) denotes bilinear upsampling to the input resolution. This approach stabilises the attribution map by smoothing layer-specific gradient noise, providing robustness against severe architectural instabilities (ResNet50V2 net ∆ = −0.098 vs. ScoreCAM ∆ = −0.612), at the cost of diluted semantic specificity in fully fine-tuned models.
4.2.2
Evaluation Protocol
4.2.3
R ES N ET 50V2 — F INE -T UNING D EGRADATION P RECEDING M ODE C OLLAPSE
All techniques show net negative E20→E30 change (Figure 3). ScoreCAM provides the clearest diagnostic signal: 0.612 (E20) → 0.014 (E25) → 0.000 (E30), detectable one full checkpoint before the AUC collapse at E30. This demonstrates C-Score’s potential as an annotation-free deployment monitoring metric for production systems.
Results
Complete result tables and visualisations are provided in Appendix A. 4.1
I NCEPTION V3 — T RANSFER -L EARNING S TASIS , F INE -T UNING R ESOLUTION
All gradient-based techniques plateau during transfer learning (GradCAM: 0.169→0.196; ScoreCAM: 0.392→0.379), reflecting gradient dispersion across parallel inception modules. EigenCAM maintains stability throughout (0.758→0.756). Fine-tuning produces a dramatic step-change but with non-monotonic GradCAM behaviour (0.875 at E25, collapsing to 0.244 at E30).
C-Score was computed at seven checkpoints per architecture (epochs 1, 5, 10, 15, 20, 25, 30) from per-architecture trajectory CSV logs. Test AUC and accuracy were extracted from epoch_metrics.csv training logs for all three architectures across all 30 epochs. The gold list was anchored to the epoch-30 reference model at τ = 0.5, yielding 317 Normal and 855 Pneumonia test-set images for binary evaluation.
4
D ENSE N ET 201 — S YSTEMATIC F INE -T UNING I MPROVEMENT
Every technique shows positive net change E20→E30. GradCAM achieves the largest absolute gain (+0.610: 0.197→0.807), consistent with the gradient flow– consistency coupling hypothesis. GradCAM++, LayerCAM, and ScoreCAM converge to 0.870–0.880 at E30, indicating method-equivalence for a well-trained stable model. EigenCAM maintains high stability throughout (≈0.635) and further improves at fine-tuning (0.846).
ℓ∈L
3.4
Global Weighted C-Score Trajectory
4.3
Full Classification Performance Trajectory
Per-Class C-Score: All Architectures
Tables 6–8 present per-class C-Score at seven checkpoints; Figure 4 visualises the trajectories.
Table 2 presents the complete per-epoch AUC and test accuracy for all three architectures across all 30 training epochs. DenseNet201 achieves monotonically increasing AUC in transfer learning (0.9184→0.9902, E1→E20), suffers boundary-reorganisation accuracy collapse at E21–23 (27–28% accuracy, AUC >0.984), recovers by E24, and peaks at E30 (AUC = 0.9945). InceptionV3 grows more slowly through transfer learning and undergoes a sharper accuracy collapse at E23 (28.75%) before resolving at E25 (96.76%). ResNet50V2 experiences catastrophic mode collapse at E23 (AUC = 0.0287) and again at E30 (AUC = 0.1034), representing a complete failure of fine-tuning stability.
4.3.1
T HE G RAD CAM C LASS G AP : D ENSE N ET 201 T RANSFER L EARNING
GradCAM on DenseNet201 produces a Pneumonia CScore that remains near-zero throughout the entire transfer learning phase (0.078 at E1, 0.007 at E5, 0.002 at E10, 0.004 at E15, 0.014 at E20) while Normal C-Score reaches 0.664 — a class gap of 0.650 at equal AUC = 0.9902. The model classifies pneumonia cases correctly but attends to entirely different regions for each patient. No classification metric — AUC, accuracy, or F1 — provides any signal of
5
Quantifying Explanation Consistency: The C-Score Metric for CAM-Based Explainability in Medical Image Classification
5.2
this class-level explanation failure. 4.4
The pre-collapse ScoreCAM signal on ResNet50V2 (CScore 0.612 → 0.014 at E25, one checkpoint before AUC collapse at E30) demonstrates that explanation consistency can deteriorate ahead of classification performance. In production systems with periodic weight updates or continued learning, C-Score monitoring provides an additional safety layer not available from AUC monitoring alone. The annotation-free nature of C-Score makes this monitoring scalable across deployment environments without requiring radiologist input at each update cycle.
The Fine-Tuning Paradox: Three Mechanisms of AUC–Consistency Dissociation
Mechanism 1 — Threshold-mediated gold-list collapse. DenseNet201 epochs 21–23: accuracy ≈ 27%, AUC > 0.984. C-Score correctly reports zero explanation consistency for the Normal class because no Normal image passes τ = 0.5. AUC conceals this operational failure entirely. Deploying at E22 (AUC = 0.9852) would mean deploying a model that generates empty explanations for the Normal class.
5.3
Mechanism 2 — Technique-specific attribution collapse at peak AUC. ScoreCAM on ResNet50V2 at E30: AUC = 0.1034 (collapsed), C-Score = 0.000. More critically, ScoreCAM C-Score degrades to 0.014 at E25 while AUC remains at 0.9902 — the consistency failure precedes the classification failure by a full checkpoint. No classification metric provides any signal of this impending collapse.
These three mechanisms collectively demonstrate that neither AUC nor global C-Score alone suffices for clinical AI quality assurance. The minimum acceptable evaluation framework is per-class C-Score tracked across the full training trajectory, using multiple CAM techniques, with explicit reporting of gold-list population at each checkpoint.
Discussion
5.1
Architecture Recommendations
Broader Implications for Clinical AI Validation
The three AUC–consistency dissociation mechanisms documented here have direct regulatory implications. The European AI Act [41] designates medical AI as high-risk and requires providers to demonstrate that high-risk AI systems produce consistent and interpretable outputs. FDA guidance [42] on AI/ML-based software as a medical device emphasises the need for performance monitoring across the model lifecycle. C-Score provides a concrete, continuously computable metric for both requirements: it quantifies explanation consistency without requiring ground-truth annotations and can be computed at any post-deployment checkpoint using only the existing test set. The analogy to interrater reliability in clinical practice [43] grounds C-Score within established clinical validation methodology.
Mechanism 3 — Class-level consistency masking in global metrics. DenseNet201 GradCAM at E20: global C-Score = 0.197, Normal = 0.664, Pneumonia = 0.014. Global aggregation masks near-total explanation failure for the clinically critical Pneumonia class. Per-class reporting is essential.
5
C-Score as a Deployment Monitoring Tool
6
Limitations
CAM scope. C-Score is formulated for 2D spatial heatmaps from convolutional layers. Not directly applicable to input-space methods (SmoothGrad [44], Integrated Gradients [45]) or model-agnostic methods (SHAP [46], LIME [47]). Layer sensitivity. C-Score values are sensitive to target layer depth. Shallower layers produce diffuse activations that may affect C-Score independently of classification quality. Systematic layer-depth sensitivity analysis is required for cross-architecture standardisation.
DenseNet201 is recommended for clinical deployment: highest stable AUC (0.9945), broadest consistency improvement across fine-tuning (average ∆ = +0.340 across six techniques), smallest class gap, and no catastrophic failures. InceptionV3 achieves marginally higher AUC (0.9949) at the cost of pronounced GradCAM instability between E25 and E30, making technique selection critical if deployed. ResNet50V2 is not recommended for clinical use: fine-tuning consistently degrades explanation consistency, ScoreCAM collapse preceded AUC collapse, and both classes register zero C-Score at E30.
Threshold dependency. The τ = 0.5 threshold determines gold-list membership. Degenerate thresholdcrossing regimes produce empty gold lists. Thresholdadaptive variants and calibrated probability weighting should be explored. Training log completeness. InceptionV3 shows transient accuracy collapse events at epochs 23 and 26 whose mechanistic origins — gradient explosion, learning-rate schedule artefacts, or batch variance — were not fully diagnosed. Future work should instrument training with perlayer gradient-norm logging. 6
Quantifying Explanation Consistency: The C-Score Metric for CAM-Based Explainability in Medical Image Classification
Single dataset. Results are from one binary classification task on one publicly available dataset. Generalisation to multi-class settings, CT/MRI modalities, and transformerbased architectures [48, 49, 50] requires dedicated investigation.
Zhu, Brandon Yang, Hershel Mehta, Tony Duan, Daisy Ding, Aarti Bagul, Curtis P Langlotz, et al. CheXNet: Radiologist-level pneumonia detection on chest x-rays with deep learning. arXiv preprint arXiv:1711.05225, 2017. doi: 10.48550/arXiv.1711. 05225.
Consistency ̸= Correctness. High C-Score certifies spatial reproducibility, not clinical correctness. A model consistently attending to spurious correlates achieves high CScore. Validation against radiologist-annotated groundtruth is necessary to confirm that consistent explanations are also clinically faithful [51].
7
[5] Alex J DeGrave, Joseph D Janizek, and Su-In Lee. AI for radiographic COVID-19 detection selects shortcuts over signal. Nature Machine Intelligence, 3:610– 619, 2021. doi: 10.1038/s42256-021-00338-7. [6] John R Zech, Marcus A Badgeley, Manway Liu, Anthony B Costa, Joseph J Titano, and Eric Karl Oermann. Variable generalization performance of a deep learning model to detect pneumonia in chest radiographs: A cross-sectional study. PLOS Medicine, 15(11):e1002683, 2018. doi: 10.1371/journal.pmed. 1002683.
Conclusion
We proposed the C-Score (Consistency Score), a confidence-weighted, annotation-free metric for quantifying intra-class explanation reproducibility of CAM-based methods in medical image classification. Evaluated across six CAM techniques, three CNN architectures, and thirty training epochs on the Kermany chest X-ray dataset, C-Score revealed three distinct mechanisms of AUC– consistency dissociation invisible to standard classification metrics. DenseNet201 demonstrated the most favourable profile for clinical deployment across both dimensions. CScore provides a practical, continuously computable quality assurance signal that complements existing evaluation frameworks and satisfies the consistency requirements emerging from regulatory guidance on clinical AI.
[7] Luke Oakden-Rayner, Jared Dunnmon, Gustavo Carneiro, and Christopher Ré. Hidden stratification causes clinically meaningful failures in machine learning for medical imaging. In Proceedings of the ACM Conference on Health, Inference, and Learning (CHIL), pages 151–159, 2020. doi: 10.1145/ 3368555.3384468. [8] Julia K Winkler, Christine Fink, Ferdinand Toberer, Alexander Enk, Teresa Deinlein, Rainer HofmannWellenhof, Luc Thomas, Aimilios Lallas, Andreas Blum, Wilhelm Stolz, et al. Association between surgical skin markings in dermoscopic images and diagnostic performance of a deep learning convolutional neural network for melanoma recognition. JAMA Dermatology, 155(10):1135–1141, 2019. doi: 10.1001/jamadermatol.2019.1735.
References [1] Varun Gulshan, Lily Peng, Marc Coram, Martin C Stumpe, Derek Wu, Arunachalam Narayanaswamy, Subhashini Venugopalan, Kasumi Widner, Tom Madams, Jorge Cuadros, et al. Development and validation of a deep learning algorithm for detection of diabetic retinopathy in retinal fundus photographs. JAMA, 316(22):2402–2410, 2016. doi: 10.1001/jama.2016.17216.
[9] Michael Roberts, Derek Driggs, Matthew Thorpe, Julian Gilbey, Michael Yeung, Stephan Ursprung, Angelica I Aviles-Rivero, Christian Etmann, Cathal McCague, Lucian Beer, et al. Common pitfalls and recommendations for using machine learning to detect and prognosticate for COVID-19 using chest radiographs and ct scans. Nature Machine Intelligence, 3:199–217, 2021. doi: 10.1038/ s42256-021-00307-0.
[2] Andre Esteva, Brett Kuprel, Roberto A Novoa, Justin Ko, Susan M Swetter, Helen M Blau, and Sebastian Thrun. Dermatologist-level classification of skin cancer with deep neural networks. Nature, 542:115–118, 2017. doi: 10.1038/nature21056. [3] Daniel S Kermany, Michael Goldbaum, Wenjia Cai, Carolina C S Valentim, Huiying Liang, Sally L Baxter, Alex McKeown, Ge Yang, Xiaokang Wu, Fangbing Yan, et al. Identifying medical diagnoses and treatable diseases by image-based deep learning. Cell, 172(5):1122–1131, 2018. doi: 10.1016/j.cell.2018. 02.010.
[10] Ziad Obermeyer, Brian Powers, Christine Vogeli, and Sendhil Mullainathan. Dissecting racial bias in an algorithm used to manage the health of populations. Science, 366(6464):447–453, 2019. doi: 10.1126/ science.aax2342. [11] Graham A McLeod, Emma A M Stanley, Tom Rosenal, and Nils D Forkert. Distinct visual biases affect humans and artificial intelligence in medical imaging
[4] Pranav Rajpurkar, Jeremy Irvin, Robyn L Ball, Kaylie 7
Quantifying Explanation Consistency: The C-Score Metric for CAM-Based Explainability in Medical Image Classification
diagnoses. npj Digital Medicine, 9(62), 2026. doi: 10.1038/s41746-025-02226-5.
[20] Ethan Barr et al. The disagreement problem in faithfulness metrics. arXiv preprint arXiv:2311.07763, 2023. doi: 10.48550/arXiv.2311.07763.
[12] Ramprasaath R Selvaraju, Michael Cogswell, Abhishek Das, Ramakrishna Vedantam, Devi Parikh, and Dhruv Batra. Grad-CAM: Visual explanations from deep networks via gradient-based localization. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), pages 618–626, 2017. doi: 10.1109/ICCV.2017.74.
[21] Aditya Chattopadhay, Anirban Sarkar, Prantik Howlader, and Vineeth N Balasubramanian. GradCAM++: Generalized gradient-based visual explanations for deep convolutional networks. In 2018 IEEE Winter Conference on Applications of Computer Vision (WACV), pages 839–847. IEEE, 2018. doi: 10.1109/WACV.2018.00097.
[13] Julius Adebayo, Justin Gilmer, Michael Muelly, Ian Goodfellow, Moritz Hardt, and Been Kim. Sanity checks for saliency maps. In Advances in Neural Information Processing Systems (NeurIPS), volume 31, pages 9525–9536, 2018.
[22] Peng-Tao Jiang, Chang-Bin Zhang, Qibin Hou, MingMing Cheng, and Yunchao Wei. LayerCAM: Exploring hierarchical class activation maps for localization. IEEE Transactions on Image Processing, 30:5875– 5888, 2021. doi: 10.1109/TIP.2021.3089943.
[14] Pieter-Jan Kindermans, Sara Hooker, Julius Adebayo, Maximilian Alber, Kristof T Schütt, Sven Dähne, Dumitru Erhan, and Been Kim. The (un)reliability of saliency methods. In Explainable AI: Interpreting, Explaining and Visualizing Deep Learning, pages 267–280. Springer, 2019. doi: 10.1007/ 978-3-030-28954-6_14.
[23] Haofan Wang, Zifan Wang, Mengnan Du, Fan Yang, Zijian Zhang, Sirui Ding, Piotr Mardziel, and Xia Hu. Score-CAM: Score-weighted visual explanations for convolutional neural networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, pages 111–119, 2020. doi: 10.1109/CVPRW50498.2020.00020.
[15] Amirata Ghorbani, Abubakar Abid, and James Zou. Interpretation of neural networks is fragile. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 33, pages 3681–3688, 2019. doi: 10.1609/aaai.v33i01.33013681.
[24] Mohammed Bany Muhammad and Mohammed Yeasin. Eigen-CAM: Class activation map using principal components. In 2020 International Joint Conference on Neural Networks (IJCNN), pages 1– 7. IEEE, 2020. doi: 10.1109/IJCNN48605.2020. 9206626.
[16] Rachel Lea Draelos and Lawrence Carin. Use HiResCAM instead of Grad-CAM for faithful explanations of convolutional neural networks. arXiv preprint arXiv:2011.08891, 2020.
[25] Siddharth Bhati, Fnu Neha, and Md Amiruzzaman. A survey on explainable artificial intelligence (XAI) techniques for visualizing deep learning models in medical imaging. Journal of Imaging, 10(10):239, 2024.
[17] Wojciech Samek, Alexander Binder, Grégoire Montavon, Sebastian Lapuschkin, and Klaus-Robert Müller. Evaluating the visualization of what a deep neural network has learned. IEEE Transactions on Neural Networks and Learning Systems, 28 (11):2660–2673, 2017. doi: 10.1109/TNNLS.2016. 2599820.
[26] Deyang Tang, Jie Chen, Liyuan Ren, Xiuqin Wang, Dan Li, and Haibing Zhang. Reviewing CAMbased deep explainable methods in healthcare. Applied Sciences, 14(10):4124, 2024. doi: 10.3390/ app14104124.
[18] Richard Tomsett, Dan Harborne, Supriyo Chakraborty, Prudhvi Gurram, and Alun Preece. Sanity checks for saliency metrics. In Proceedings of the AAAI Conference on Artificial Intelligence, pages 6021–6029, 2020. doi: 10.1609/aaai.v34i04.6064.
[27] Bas H M van der Velden, Hugo J Kuijf, Kenneth G A Gilhuijs, and Max A Viergever. Explainable artificial intelligence (XAI) in deep learning-based medical image analysis. Medical Image Analysis, 79: 102470, 2022. doi: 10.1016/j.media.2022.102470.
[19] Satyapriya Krishna, Tessa Han, Alex Gu, Javin Pombra, Shahin Jabbari, Steven Wu, and Himabindu Lakkaraju. The disagreement problem in explainable machine learning: A practitioner’s perspective. Transactions on Machine Learning Research, 2024. doi: 10.48550/arXiv.2202.01602. arXiv:2202.01602 (2022); TMLR publication (2024).
[28] Suman Suara et al. Is grad-CAM explainable in medical images? arXiv preprint arXiv:2307.10506, 2023. doi: 10.48550/arXiv.2307.10506. [29] Bolei Zhou, Aditya Khosla, Agata Lapedriza, Aude Oliva, and Antonio Torralba. Learning deep features 8
Quantifying Explanation Consistency: The C-Score Metric for CAM-Based Explainability in Medical Image Classification
for discriminative localization. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2921–2929, 2016. doi: 10.1109/CVPR.2016.319.
[38] Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Identity mappings in deep residual networks. In European Conference on Computer Vision (ECCV), volume 9908 of Lecture Notes in Computer Science, pages 630–645. Springer, 2016. doi: 10.1007/978-3-319-46493-0_38.
[30] Junhee Lee, Hyeonseong Cho, Yun Jang Pyun, SukJu Kang, and Hyoungsik Nam. Heatmap assisted accuracy score evaluation method for machine-centric explainable deep neural networks. IEEE Access, 10: 64832–64849, 2022. doi: 10.1109/ACCESS.2022. 3184453.
[39] Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Fei-Fei Li. ImageNet: A large-scale hierarchical image database. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 248–255, 2009. doi: 10.1109/ CVPR.2009.5206848.
[31] Sara Hooker, Dumitru Erhan, Pieter-Jan Kindermans, and Been Kim. A benchmark for interpretability methods in deep neural networks. In Advances in Neural Information Processing Systems (NeurIPS), 2019. doi: 10.48550/arXiv.1806.10758.
[40] Diederik P Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In International Conference on Learning Representations (ICLR), 2015. doi: 10.48550/arXiv.1412.6980.
[32] Anna Hedström, Leander Weber, Daniel Krakowczyk, Dilyara Bareeva, Franz Motzkus, Wojciech Samek, Sebastian Lapuschkin, and Marina M-C Müller. Quantus: An explainable AI toolkit for responsible evaluation of neural network explanations and beyond. Journal of Machine Learning Research, 24(34):1–11, 2023.
[41] European Parliament and Council of the European Union. Regulation on artificial intelligence (AI Act). Technical Report EU 2024/1689, Official Journal of the European Union, 2024. [42] U.S. Food and Drug Administration. AI/ML-based software as a medical device (SaMD) action plan. Technical report, U.S. Food and Drug Administration, 2021.
[33] Anna Hedström, Philine Bommer, Kristoffer K Wickström, Wojciech Samek, Sebastian Lapuschkin, and Marina M-C Müller. The meta-evaluation problem in explainable AI: Identifying reliable estimators with MetaQuantus. Transactions on Machine Learning Research, 2023. Featured Certification.
[43] J Richard Landis and Gary G Koch. The measurement of observer agreement for categorical data. Biometrics, 33(1):159–174, 1977. doi: 10.2307/2529310.
[34] Cemre Ozer et al. Consistent explainable image quality assessment for medical imaging. Health Information Science and Systems, 14:31, 2025. doi: 10.1007/s13755-025-00411-0.
[44] Daniel Smilkov, Nikhil Thorat, Been Kim, Fernanda Viégas, and Martin Wattenberg. SmoothGrad: Removing noise by adding noise. arXiv preprint arXiv:1706.03825, 2017. doi: 10.48550/arXiv.1706. 03825.
[35] Miguel A. Lago, Ghada Zamzmi, Brandon Eich, and Jana G. Delfino. Evaluating explainability: A framework for systematic assessment and reporting of explainable ai features, 2025. URL https:// arxiv.org/abs/2506.13917.
[45] Mukund Sundararajan, Ankur Taly, and Qiqi Yan. Axiomatic attribution for deep networks. In Proceedings of the 34th International Conference on Machine Learning (ICML), pages 3319–3328, 2017. doi: 10.48550/arXiv.1703.01365.
[36] Gao Huang, Zhuang Liu, Laurens van der Maaten, and Kilian Q Weinberger. Densely connected convolutional networks. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2261–2269, 2017.
[46] Scott M Lundberg and Su-In Lee. A unified approach to interpreting model predictions. In Advances in Neural Information Processing Systems (NeurIPS), pages 4765–4774, 2017. doi: 10.48550/arXiv.1705. 07874.
[37] Christian Szegedy, Vincent Vanhoucke, Sergey Ioffe, Jon Shlens, and Zbigniew Wojna. Rethinking the inception architecture for computer vision. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR), pages 2818–2826, 2016. doi: 10.1109/CVPR.2016.308.
[47] Marco Tulio Ribeiro, Sameer Singh, and Carlos Guestrin. “Why Should I Trust You?”: Explaining the predictions of any classifier. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, pages 1135– 1144, 2016. doi: 10.1145/2939672.2939778. 9
Quantifying Explanation Consistency: The C-Score Metric for CAM-Based Explainability in Medical Image Classification
[48] Mingxing Tan and Quoc Le. EfficientNet: Rethinking model scaling for convolutional neural networks. In Proceedings of the 36th International Conference on Machine Learning (ICML), pages 6105–6114, 2019. doi: 10.48550/arXiv.1905.11946. [49] Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A ConvNet for the 2020s. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11976–11986, 2022. [50] Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16×16 words: Transformers for image recognition at scale. In International Conference on Learning Representations (ICLR), 2021. doi: 10.48550/arXiv.2010.11929. [51] Agustina J Larrazabal, Nicolás Nieto, Victoria Peterson, Diego H Milone, and Enzo Ferrante. Gender imbalance in medical imaging datasets produces biased classifiers for computer-aided diagnosis. Proceedings of the National Academy of Sciences, 117(23):12592– 12594, 2020. doi: 10.1073/pnas.1919012117.
10
Quantifying Explanation Consistency: The C-Score Metric for CAM-Based Explainability in Medical Image Classification
A
Result Tables and Figures
Table 2. Complete 30-epoch test AUC and accuracy for all three architectures. TL = Transfer Learning (epochs 1–20); FT = FineTuning (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. DenseNet201
InceptionV3
ResNet50V2
Ep.
Phase
AUC
Acc.
AUC
Acc.
AUC
Acc.
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20
TL TL TL TL TL TL TL TL TL TL TL TL TL TL TL TL TL TL TL TL
0.9184 0.9656 0.9657 0.9804 0.9848 0.9855 0.9859 0.9865 0.9882 0.9886 0.9889 0.9890 0.9893 0.9893 0.9903 0.9890 0.9901 0.9902 0.9899 0.9902
61.60% 84.64% 45.48% 90.96% 94.03% 92.58% 94.97% 90.44% 94.80% 94.88% 94.88% 95.14% 95.31% 94.80% 90.79% 95.73% 95.73% 95.90% 95.14% 94.80%
0.8705 0.9310 0.9509 0.9542 0.9592 0.9626 0.9602 0.9641 0.9633 0.9664 0.9673 0.9672 0.9675 0.9661 0.9679 0.9661 0.9678 0.9676 0.9684 0.9680
66.13% 64.33% 69.97% 86.35% 78.33% 80.46% 85.49% 84.13% 87.63% 86.86% 87.20% 87.03% 88.99% 88.23% 88.91% 89.25% 87.88% 87.46% 88.82% 89.33%
0.9582 0.9760 0.9807 0.9800 0.9876 0.9885 0.9859 0.9888 0.9894 0.9892 0.9908 0.9909 0.9895 0.9900 0.9902 0.9893 0.9901 0.9891 0.9888 0.9898
81.48% 94.20% 82.42% 94.71% 95.82% 95.56% 95.82% 90.79% 94.71% 93.52% 95.05% 94.28% 95.90% 84.39% 94.97% 92.06% 92.49% 94.62% 95.14% 94.20%
21 22 23 24 25 26 27 28 29 30
FT FT FT FT FT FT FT FT FT FT
0.9842 0.9852 0.9891 0.9910 0.9867 0.9844 0.9931 0.9927 0.9941 0.9945
27.73% 27.05% 27.05% 83.70% 72.95% 72.95% 96.93% 96.76% 86.77% 94.20%
0.9297 0.9648 0.9836 0.9892 0.9902 0.9930 0.9925 0.9949 0.9943 0.9949
84.30% 72.95% 28.75% 69.20% 96.76% 38.23% 87.46% 97.61% 94.62% 94.62%
0.9876 0.9807 0.0287 0.9885 0.9902 0.9868 0.9933 0.9938 0.9947 0.1034
79.86% 42.49% 70.99% 44.45% 78.75% 82.34% 95.39% 74.91% 97.01% 72.95%
Table 3. Global weighted C-Score for DenseNet201. Green : C-Score ≥ 0.80; amber : < 0.30; red : 0.000. Step-change improvement at TL→FT transition observable across all methods. Technique
E1
E5
E10
E15
E20 (TL End)
E25 (Mid-FT)
E30 (FT End)
GradCAM GradCAM++ LayerCAM ScoreCAM EigenCAM MS-GradCAM++
0.113 0.358 0.403 0.460 0.635 0.322
0.168 0.461 0.479 0.563 0.634 0.380
0.170 0.546 0.569 0.622 0.635 0.439
0.198 0.566 0.583 0.632 0.636 0.460
0.197 0.549 0.559 0.630 0.635 0.445
0.744 0.916 0.915 0.933 0.908 0.644
0.807 0.870 0.871 0.880 0.846 0.618
∆ (E30−E20): +0.610 (GradCAM); +0.321 (GradCAM++); +0.312 (LayerCAM); +0.250 (ScoreCAM); +0.211 (EigenCAM); avg: +0.267
11
Quantifying Explanation Consistency: The C-Score Metric for CAM-Based Explainability in Medical Image Classification
Table 4. Global weighted C-Score for InceptionV3. Technique
E1
E5
E10
E15
E20 (TL End)
E25 (Mid-FT)
E30 (FT End)
GradCAM GradCAM++ LayerCAM ScoreCAM EigenCAM MS-GradCAM++
0.169 0.475 0.567 0.392 0.758 0.419
0.242 0.508 0.583 0.386 0.759 0.417
0.209 0.492 0.572 0.383 0.758 0.417
0.195 0.484 0.567 0.381 0.757 0.415
0.196 0.485 0.568 0.379 0.756 0.419
0.875 0.808 0.802 0.790 0.896 0.659
0.244 0.762 0.763 0.759 0.852 0.654
∆ (E30−E20): +0.048 (GradCAM); +0.277 (GradCAM++); +0.195 (LayerCAM); +0.380 (ScoreCAM); +0.096 (EigenCAM); avg: +0.158
Table 5. Global weighted C-Score for ResNet50V2. Net E20→E30 change negative for majority of techniques. ScoreCAM trajectory 0.612 → 0.014 → 0.000 provides pre-collapse early-warning signal one checkpoint before AUC collapse at E30. Technique
E1
E5
E10
E15
E20 (TL End)
E25 (Mid-FT)
E30 (FT End)
GradCAM GradCAM++ LayerCAM ScoreCAM EigenCAM MS-GradCAM++
0.422 0.320 0.513 0.517 0.589 0.313
0.387 0.613 0.667 0.698 0.685 0.508
0.573 0.642 0.681 0.629 0.693 0.527
0.400 0.607 0.662 0.621 0.692 0.508
0.385 0.593 0.654 0.612 0.688 0.507
0.272 0.676 0.675 0.014 0.678 0.533
0.370 0.478 0.489 0.000 0.495 0.409
∆ (E30−E20): −0.015 (GradCAM); −0.115 (GradCAM++); −0.165 (LayerCAM); −0.612 (ScoreCAM); −0.193 (EigenCAM); avg: −0.162
Figure 1. Global weighted C-Score trajectory across training phases.
12
Quantifying Explanation Consistency: The C-Score Metric for CAM-Based Explainability in Medical Image Classification
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.
13
Quantifying Explanation Consistency: The C-Score Metric for CAM-Based Explainability in Medical Image Classification
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 τ = 0.5 at this checkpoint during boundary reorganisation. E1
E5
E10
E15
E20 (TL)
E25 (FT)
E30 (FT)
0.159 0.424 0.471 0.539 0.680 0.370
0.593 0.669 0.672 0.621 0.682 0.504
0.606 0.694 0.696 0.688 0.688 0.535
0.663 0.728 0.729 0.738 0.689 0.586
0.664 0.718 0.716 0.714 0.688 0.573
0.000 0.000 0.000 0.000 0.000 0.000
0.924 0.918 0.919 0.907 0.895 0.674
0.007 0.382 0.406 0.541 0.615 0.333
0.002 0.490 0.520 0.597 0.615 0.402
0.004 0.498 0.522 0.587 0.614 0.408
0.014 0.483 0.498 0.597 0.614 0.395
0.744 0.916 0.915 0.933 0.908 0.644
0.761 0.851 0.852 0.869 0.826 0.596
Technique Normal (Class 0) GradCAM GradCAM++ LayerCAM ScoreCAM EigenCAM MS-GradCAM++
Pneumonia (Class 1) GradCAM GradCAM++ LayerCAM ScoreCAM EigenCAM MS-GradCAM++
0.078 0.310 0.352 0.401 0.603 0.286
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. E1
E5
E10
E15
E20 (TL)
E25 (FT)
E30 (FT)
0.047 0.469 0.613 0.292 0.770 0.393
0.288 0.542 0.642 0.332 0.777 0.407
0.380 0.536 0.631 0.303 0.775 0.407
0.317 0.516 0.620 0.286 0.774 0.400
0.335 0.509 0.623 0.286 0.774 0.396
0.840 0.872 0.866 0.845 0.938 0.753
0.847 0.851 0.852 0.852 0.922 0.759
0.218 0.491 0.553 0.414 0.750 0.422
0.138 0.474 0.548 0.416 0.750 0.421
0.146 0.471 0.546 0.419 0.750 0.421
0.140 0.475 0.546 0.417 0.749 0.429
0.887 0.785 0.779 0.770 0.881 0.626
0.008 0.728 0.728 0.723 0.825 0.613
Technique Normal (Class 0) GradCAM GradCAM++ LayerCAM ScoreCAM EigenCAM MS-GradCAM++
Pneumonia (Class 1) GradCAM GradCAM++ LayerCAM ScoreCAM EigenCAM MS-GradCAM++
0.244 0.479 0.539 0.453 0.750 0.435
14
Quantifying Explanation Consistency: The C-Score Metric for CAM-Based Explainability in Medical Image Classification
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. E1
E5
E10
E15
E20 (TL)
E25 (FT)
E30 (FT)
0.333 0.364 0.517 0.508 0.593 0.333
0.095 0.656 0.715 0.882 0.745 0.495
0.573 0.662 0.711 0.652 0.727 0.518
0.455 0.643 0.696 0.645 0.728 0.508
0.481 0.634 0.688 0.649 0.723 0.508
0.798 0.798 0.798 0.041 0.802 0.595
0.000 0.000 0.000 0.000 0.000 0.000
0.493 0.598 0.650 0.632 0.664 0.512
0.572 0.635 0.669 0.621 0.680 0.530
0.379 0.593 0.648 0.611 0.678 0.507
0.348 0.578 0.640 0.597 0.675 0.507
0.000 0.613 0.611 0.000 0.614 0.500
0.370 0.478 0.489 0.000 0.495 0.409
Technique Normal (Class 0) GradCAM GradCAM++ LayerCAM ScoreCAM EigenCAM MS-GradCAM++
Pneumonia (Class 1) GradCAM GradCAM++ LayerCAM ScoreCAM EigenCAM MS-GradCAM++
0.463 0.300 0.511 0.521 0.587 0.303
Figure 4. Per-class C-Score trajectory by architecture.
15