Answers are grounded in this paper and cite their sources.
On the MSID dataset, which includes four classes—monkeypox, chickenpox, measles, and normal—to introduce MLF-Net, a small hybrid CNN-Transformer that fuses features focused on lesions to recognize monkeypox in multiple classes. From class-aware training (using effective-number weights, label smoothing, and focal loss) to data-level regularization (using MixUp/CutMix and mild illumination normalization) and architectural priors (using spatial/channel attention, gated CNN-ViT fusion, and lesion-aware pooling), the pipeline achieves it all. Probabilities are calibrated for trustworthy triage using post-hoc temperature scaling. Accuracy 93.68%, Macro-F1 93.60%, Macro-AUROC 98.10%, Macro-AUPRC 96.50%, Brier 0.059, and ECE 0.024 (post-temp) are achieved by MLF-Net on a stratified MSID split. There is a high level of sensitivity to monkeypox (Recall 93.48%, F1 93.99%) and a clear demarcation of normal (F1 96.17%), but most of the remaining confusions are related to chickenpox and measles. The component utility is confirmed by ablations: 2.9 percentage points for Macro-F1 is reduced when lesion attention is removed and 2.1 percentage points when hybrid fusion is disabled; ECE remains at 0.045 (+0.021) after calibration are disabled. Research on robustness has shown that external domain performance is partially restored after temperature recalibration (Macro-F1 +0.8 pp compared to non-recalibrated), but graceful degradation occurs under noise, blur, and brightness (e.g., severity-5 Macro-F1 drop ≈ 4.7 pp). The efficiency profiling shows that it can be deployed on common hardware with around 14.2 million parameters, 14 milliseconds per image for the GPU, and 180 milliseconds per image for the HP i7 CPU. Applying TTA ×5 (five-view Test-Time Augmentation) yields approximately +0.5pp Macro-F1 at the cost of increased inference latency. Clear trade-offs between accuracy, robustness, and inference budget are highlighted by these results, which also show that MLF-Net is prepared for calibrated Mpox screening with limited resources.
No takes yet. Share an insight, caveat, or question.
V et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: