Table 3 The comparison with the standard models.
From: A hybrid parallel convolutional spiking neural network for enhanced skin cancer detection
Model | Accuracy | Sensitivity | Specificity | F1-score | Training time | Inference time | Advantages | Limitations |
|---|---|---|---|---|---|---|---|---|
PCSN-Net (Proposed) | 0.957 | 0.947 | 0.926 | 0.927 | \(\tilde{2}\) h (on single GPU) | \(\tilde{2}0\) ms per image | High accuracy, energy-efficient | Computation-ally intensive |
EfficientNet | 0.935 | 0.92 | 0.91 | 0.915 | \(\tilde{1}.5\) h | \(\tilde{1}5\) ms per image | Efficient, low computational cost | May require fine-tuning on medical images |
DenseNet | 0.94 | 0.93 | 0.915 | 0.922 | \(\tilde{2}.5\) h | \(\tilde{3}0\) ms per image | Fast convergence, high accuracy | Increased memory requirement |
ResNet + CBAM | 0.945 | 0.94 | 0.92 | 0.93 | \(\tilde{3}\) h | \(\tilde{2}5\) ms per image | Enhanced focus on relevant features | Higher computational cost |
Xception | 0.938 | 0.925 | 0.918 | 0.92 | \(\tilde{2}\) h | \(\tilde{2}2\) ms per image | Optimized for image processing | Requires extensive data augmentation |
Inception-ResNet-V2 | 0.95 | 0.935 | 0.925 | 0.932 | \(\tilde{3}.5\) h | \(\tilde{2}8\) ms per image | High accuracy, robust for complex tasks | High computational complexity |
AlexNet | 0.89 | 0.875 | 0.87 | 0.872 | Low (2–3 h on GPU) | Very Low (40 ms) | Simple, fast training, good baseline | Limited depth, lower accuracy |
VGG-16 | 0.91 | 0.9 | 0.898 | 0.902 | High (12–15 h on GPU) | Moderate (120 ms) | High accuracy, well-known baseline | Large model size, high computation |
ResNet-50 | 0.925 | 0.915 | 0.913 | 0.917 | Moderate (6–8 h on GPU) | Moderate (100 ms) | Good balance of depth and accuracy | Requires more memory and tuning |
