Classification of skin cancer from images of moles

Model Descriptions:

Model 1: Basic Model

Structure:

• Normalization layer
• Three convolutional blocks each with:
  • Conv2D layer with 32 filters of size 3x3 and ReLU activation
  • MaxPooling2D layer
• Flatten layer
• Dense layer with 128 units and ReLU activation
• Dense layer with a number of units equal to num_classes.

Characteristics:

• A simple convolutional neural network (CNN) designed as a baseline for image classification tasks.
• Uses ReLU as the activation function which is widely used due to its computational efficiency.
• The final dense layer determines class probabilities.

Model 2: Base + Data Augmentation

Structure:

• Data Augmentation layer
• [All layers from the Basic Model]

Characteristics:

• Builds upon the Basic Model by introducing a data augmentation layer at the beginning.
• Data augmentation helps artificially increase the size of the training dataset by applying random, realistic modifications to the training images, aiding in improving model generalization.
• Common augmentations include random rotations, zooms, shifts, and flips.

Model 3: Base + Data Augmentation + Additional Layers

Structure:

• [All layers from Model 2]
• Conv2D layer with 32 filters of size 3x3 without activation
• LeakyReLU activation layer
• Conv2D layer with 32 filters of size 3x3 without activation
• BatchNormalization layer
• LeakyReLU activation layer
• Conv2D layer with 64 filters of size 5x5 with ReLU activation
• GlobalAveragePooling2D layer

Characteristics:

• This model further builds upon Model 2 by introducing more convolutional layers and a global average pooling layer.
• The LeakyReLU activation function is introduced, which allows a small gradient when the unit is not active, potentially helping with the vanishing gradient problem in deep networks.
• Batch normalization is added to improve the training speed, reduce the dependency on the initialization, and act as a form of regularization, potentially preventing overfitting.