Research
XDenseQNet
Quantum-enhanced, explainable DenseNet for monkeypox skin lesion classification. Beats every classical baseline tested.
XDenseQNet classifies monkeypox skin lesions across 4 classes on the MSID dataset. The architecture is a DenseNet121 backbone with CBAM attention, feeding a parameterised quantum circuit of 4 qubits and 2 layers as the classification head.
The result outperforms every classical baseline it was compared against, including ConvNeXt-Tiny, EfficientNet-B0, and Swin-Tiny, while running inference in under 10 milliseconds.
Results
- Accuracy
- 95.83%
- F1 score
- 0.9470
- Precision
- 0.9468
- Recall
- 0.9531
- ROC-AUC
- 0.9945
- Inference time
- 9.85 ms
Against classical baselines
- XDenseQNet: 95.83%
- ConvNeXt-Tiny: 94.17%
- EfficientNet-B0: 92.50%
- MobileNet-V2: 92.50%
- ResNet50: 90.83%
- Swin-Tiny: 90.00%
- ResNet18: 88.33%
- VGG16: 87.50%
Ablation
Removing the quantum head drops accuracy from 95.83% to 90.00%. Removing both CBAM and the quantum head gives 88.33%. Swapping the head for an SVM with an RBF kernel gives 82.50%. The quantum circuit is doing real work rather than decorating the architecture.
Explainability
The X in the name is the explainability component. The model exposes attention and attribution maps so a prediction can be inspected against the lesion it was made from, which is the minimum bar for anything medical.
Stack
- Quantum
- PennyLane 0.36+
- Deep learning
- PyTorch 2.1+DenseNet121CBAM
- Data
- MSID dataset