Table 2 Comparison of CNN and Transformer-based models in terms of parameter configuration, computational requirements, and structural complexity for skin disease image analysis.
Method | Batch size | Loss function | Optimizer | Learning rate | Parameters | GFLOPs |
|---|---|---|---|---|---|---|
AlexNet (2012) | 4 | Cross entropy loss | SGD | 0.001 | 57.01 M | 1.42 |
VGG19 (2014) | 4 | Cross entropy loss | SGD | 0.001 | 139.58 M | 39.28 |
GoogLeNet (2015) | 4 | Cross entropy loss | SGD | 0.001 | 5.60 M | 3.00 |
SqueezeNet (2016) | 4 | Cross entropy loss | SGD | 0.001 | 0.73 M | 1.47 |
ResNet-50 (2016) | 4 | Cross entropy loss | SGD | 0.001 | 23.51 M | 8.18 |
DenseNet-121 (2017) | 4 | Cross entropy loss | SGD | 0.001 | 6.95 M | 5.66 |
MobileNetV3 (2019) | 4 | Cross entropy loss | SGD | 0.001 | 1.52 M | 0.11 |
EfficientNetV2 (2021) | 4 | Cross entropy loss | SGD | 0.001 | 20.18 M | 5.70 |
ViT (2020) | 4 | Cross entropy loss | SGD | 0.001 | 85.80 M | 24.04 |
Swin (2021) | 4 | Cross entropy loss | SGD | 0.001 | 86.74 M | 21.10 |
CvT (2021) | 4 | Cross entropy loss | SGD | 0.001 | 19.61 M | 8.18 |
DaViT (2022) | 4 | Cross entropy loss | SGD | 0.001 | 86.93 M | 30.56 |
MaxViT (2022) | 4 | Cross entropy loss | SGD | 0.001 | 30.40 M | 10.96 |
GC ViT (2023) | 4 | Cross entropy loss | SGD | 0.001 | 89.29 M | 27.78 |
FastViT-S12 (2023) | 4 | Cross entropy loss | SGD | 0.001 | 8.45 M | 2.80 |
SHViT-S1 (2024) | 4 | Cross entropy loss | SGD | 0.001 | 13.79 M | 1.21 |
