COVID-QU-Ex: CXR Benchmark for COVID Diagnosis

• COVID-QU-Ex is a chest X-ray dataset with 85,318 images spanning COVID-19, Non-COVID Pneumonia, and Normal, serving as a benchmark for diagnostic accuracy and safety auditing.
• The dataset supports efficient classification via knowledge distillation, where high-capacity teacher models guide a compact student model for interpretable and resource-friendly performance.
• It facilitates adversarial robustness testing with spatial attention transfer, highlighting vulnerabilities in standard defenses and informing deployment in low-resource clinical environments.

COVID-QU-Ex is a chest X-ray (CXR) dataset used for multi-class COVID-19 image classification, documented in later arXiv studies as containing 85,318 chest X-ray images spanning COVID-19, Non-COVID Pneumonia, and Normal classes, and serving as a benchmark for both robustness auditing and efficient, explainable model compression (Gabriel et al., 16 May 2026, Mir et al., 21 Aug 2025). Within the published record available here, it appears less as a single fixed protocol than as a research substrate for two distinct lines of work: safety analysis under adversarial stress and deployment-oriented classification with knowledge distillation. The resulting literature positions COVID-QU-Ex at the intersection of radiographic diagnosis, model efficiency, explainability, and clinical reliability.

1. Dataset scope and documented class structure

COVID-QU-Ex is described as a chest X-ray dataset with 85,318 images and three diagnostic categories: COVID-19, Non-COVID Pneumonia, and Normal (Gabriel et al., 16 May 2026). In a separate classification study, the same dataset is treated as a benchmark for multiple classes “including COVID-19, healthy, non-COVID pneumonia,” which indicates a label vocabulary shift between “Normal” and “healthy” without changing the underlying three-way diagnostic framing (Mir et al., 21 Aug 2025).

The most explicit experimental sampling reported for COVID-QU-Ex comes from a robustness audit that used 150 images from the Test split (50/class) and a DenseNet121 model fine-tuned on 600 training images (200/class) (Gabriel et al., 16 May 2026). This does not define the full dataset protocol, but it shows that researchers used controlled class-balanced subsets for safety evaluation.

This suggests that COVID-QU-Ex is valuable not only because of nominal scale, but because it supports sharply defined experimental slices for stress-testing and comparative benchmarking.

2. Role in efficient classification and knowledge distillation

A later study uses COVID-QU-Ex as one of two benchmark datasets in an explainable knowledge distillation pipeline for medical image classification (Mir et al., 21 Aug 2025). In that framework, high-capacity teacher models—including VGG19, Visformer-S, AutoFormer-V2-T, and DenseNet-201—supervise a compact OFA-595 student model. The stated objective is efficient and interpretable COVID-19 diagnosis from CXR images under hybrid supervision, combining hard labels with teacher soft targets.

The distillation objective is given in two forms. The general form is

$$L_{KD} = \alpha \cdot L_{CE}(y, \hat{y}_s) + (1-\alpha) \cdot L_{KL}(\mathrm{soft}(y_t), \mathrm{soft}(\hat{y}_s)),$$

and the final hybrid loss used is

$$L_{KD} = \alpha \cdot L_{FBCE}(p_s, y) + (1-\alpha) \cdot L_{MSE}(p_s, p_t),$$

with $\alpha = 0.5$ (Mir et al., 21 Aug 2025).

The reported COVID-QU-Ex results for distilled OFA-595 are as follows:

Teacher	Accuracy (%)	F1-score (%)
VGG-19	94.3	94.1
DenseNet-201	96.8	96.2
OFA-595	96.7	97.3
Visformer-S	97.0	97.1
EEEA-Net-C2	96.9	98.7
AutoFormer	97.0	97.5

The corresponding AUC values range from 86.1% to 93.7% (Mir et al., 21 Aug 2025). The paper emphasizes that transformer-based teachers such as Visformer-S and AutoFormer-V2-T yielded the highest accuracy and F1-score on COVID-QU-Ex, while the student model retained strong predictive performance with significantly reduced parameters and inference time. The student is explicitly described as being selected for Pareto efficiency and for suitability in resource-constrained clinical environments.

3. Explainability and spatial attention transfer

COVID-QU-Ex has also been used in conjunction with Score-CAM-based visualizations to inspect the spatial focus of both teacher and student models (Mir et al., 21 Aug 2025). The reported finding is that student heatmaps strongly align with those of teacher models, indicating that distillation transfers not only class decisions but also salient spatial emphasis.

The same study states that, in COVID-19 and non-COVID pneumonia radiographs, teacher models attend to broader lung fields, whereas student models often learn more localized, precise attention (Mir et al., 21 Aug 2025). The qualitative interpretation offered there is that the student is not merely reproducing logits; it is inheriting a clinically relevant attention pattern. Figure 1, as described in the paper, presents the original image alongside teacher and student Score-CAM heatmaps as visual evidence of this alignment.

This use of COVID-QU-Ex situates the dataset within explainable medical AI rather than within pure classification benchmarking alone.

4. Adversarial fragility and diagnostic collapse

A systematic safety audit uses COVID-QU-Ex to evaluate adversarial image fragility in a DenseNet121 model, the architecture underlying CheXNet (Gabriel et al., 16 May 2026). The perturbation method is the Fast Gradient Method (FGM), written as

$$x_{\mathrm{adv}} = x + \epsilon \cdot \mathrm{sign}(\nabla_x J(\theta, x, y)).$$

At the first attack point, $\epsilon = 0.021$, corresponding to a 2.1% per-pixel intensity change described as imperceptible, diagnostic accuracy drops from 89.3% on clean inputs to 62.0% with adversarial perturbation, with 95% CI: 54.2–69.8% (Gabriel et al., 16 May 2026). The paper characterizes this as a –27.3% decline and notes that higher $\epsilon$ values drive performance further downward, reaching 43.3% at $\epsilon = 0.150$ and approaching random chance at sufficiently high perturbation.

The per-class breakdown at $\epsilon = 0.021$ is especially notable:

• Normal: 78.0% → 42.0%
• COVID-19: 100.0% → 72.0%
• Non-COVID Pneumonia: 90.0% → 70.0%

The most severe degradation occurs in the Normal class, and the paper specifically notes misclassification patterns in which many Normal patients are diagnosed as pneumonia (Gabriel et al., 16 May 2026). A common misconception in benchmark-centered reporting is that strong clean accuracy implies operational safety. COVID-QU-Ex-based adversarial evidence directly contradicts that assumption.

5. Failure of standard defenses

The same audit evaluates three standard defensive strategies at $\epsilon = 0.021$ and concludes that none restored diagnostic safety (Gabriel et al., 16 May 2026). The reported outcomes are:

• Gaussian Smoothing with $\sigma = 1.0$: 34.7% accuracy
• Ensemble Voting across 5 geometric augmentations: 72.7% accuracy
• Mini adversarial training: quantitative recovery not specified in the summary, but described as insufficient to restore safety

The Gaussian smoothing result is especially important because it falls below the no-defense adversarial accuracy and is described as contradicting naive expectations about preprocessing as a defense. The stated explanation is that blurring removes diagnostically critical texture while failing to eliminate the adversarial signal (Gabriel et al., 16 May 2026).

A plausible implication is that COVID-QU-Ex is not only a classification benchmark but also a useful substrate for mapping the limits of conventional robustness heuristics in radiographic AI.

6. Deployment context, low-resource settings, and future directions

The adversarial audit using COVID-QU-Ex is embedded in a broader study of clinical AI in low-resource healthcare settings, specifically Primary Health Centre (PHC) deployment in Nigeria (Gabriel et al., 16 May 2026). Although the companion language experiments operate on text case vignettes rather than on COVID-QU-Ex itself, they establish the deployment environment in which image models may be combined with multilingual interfaces or LLM-mediated decision support. In that setting, Llama3.1:8b drops from 80.0% in Standard English to 65.0% in Nigerian Pidgin and 60.0% in Yoruba-inflected English, while NatLAS drops from 85.0% to 55.0% in Pidgin, with 50% diagnosis consistency there (Gabriel et al., 16 May 2026).

This does not modify the dataset’s image content, but it narrows the interpretive gap between benchmark success and clinical readiness. The safety paper therefore argues that models should publish an Adversarial Robustness Map, be evaluated on community-level language variants, and undergo adversarial and cross-lingual audit prior to deployment (Gabriel et al., 16 May 2026).

In parallel, the distillation study frames COVID-QU-Ex as a basis for practical, trustworthy medical AI solutions, emphasizing resource-limited clinics, edge deployment, and future work targeting sub-100ms latency, broader clinical testing, regulatory compliance, and PACS integration (Mir et al., 21 Aug 2025). Taken together, these studies portray COVID-QU-Ex not simply as a static CXR corpus, but as a benchmark through which accuracy, efficiency, explainability, and failure envelopes can be studied in a clinically consequential way.