Multiclass Skin Lesion Classification Using Capsule Network with Dynamic Routing

Accurate classification is essential for identifying the skin lesions. This paper proposes a CapsNet architecture for classifying lesions into seven groups using the HAM10000 dataset. Images were resized to 28 × 28 × 28 pixels, pixel values were normalized, and the data was divided into 80% training and 20% validation groups as part of the preprocessing step. In order to efficiently capture spatial hierarchies in the data, the CapsNet design makes use of dynamic routing. It includes a decoder network for input reconstruction, which helps with regularization, a convolutional layer for feature extraction, and Primary and Digit Capsule Layers for class-specific representation. The model demonstrated CapsNet's ability to maintain spatial information and dependable classification accuracy.