Table 1.

Comparison of different methods.

Technical category	Methods	Advantages	Limitations	Relevance to this Study
ML	SVM+PCA and AdaBoost	Small parameter count, suitable for edge deployment	Relying on manual features	Highlighting the necessity of automatic feature learning
Basic CNN	VGG and AlexNet	End to end training with high accuracy	High computational cost	Comparison baseline for lightweight design
Lightweight CNN	MobileNetV3 and EfficientNet	Balancing efficiency and accuracy	Poor adaptability to geometric deformation of lesions	Improved target for deformable convolution
Attention Model	Vision Transformer	Dependency modeling	Difficulty in edge deployment	Global attention hierarchical design