Skip to main content
The skin cancer detection model uses a custom VGG-16 inspired Convolutional Neural Network (CNN) architecture, specifically designed to classify dermatological images into seven distinct categories of skin lesions.

​
Model overview

The model is a Sequential architecture built with Keras and converted to TensorFlow.js format for web deployment. It accepts RGB images of size 75×100 pixels (3 channels) and outputs probability distributions across 7 classification categories.
The model is distributed as a model.json file along with 25 binary weight shards (group1-shard1of25.bin through group1-shard25of25.bin), totaling approximately 100MB.

​
Input specifications

  • Input shape: [75, 100, 3] (height × width × channels)
  • Data type: float32
  • Color space: RGB
  • Preprocessing: Images are normalized to float32 values

​
Architecture layers

The network consists of 27 layers organized into three main stages: convolutional feature extraction, dimensionality reduction, and classification.

​
Stage 1: Initial convolutional blocks

The first stage applies six consecutive 3×3 convolutional layers with 64 filters each, using same padding to preserve spatial dimensions. 
Layer 0: InputLayer - Input shape [null, 75, 100, 3]
Layer 1: Conv2D - 64 filters, kernel 3×3, padding same
Layer 2: Activation - ReLU
Layer 3: Conv2D - 64 filters, kernel 3×3, padding same
Layer 4: Activation - ReLU
Layer 5: Conv2D - 64 filters, kernel 3×3, padding same
Layer 6: Activation - ReLU
Layer 7: Conv2D - 64 filters, kernel 3×3, padding same
Layer 8: Activation - ReLU
Layer 9: Conv2D - 64 filters, kernel 3×3, padding same
Layer 10: Activation - ReLU
Layer 11: Conv2D - 64 filters, kernel 3×3, padding valid
Layer 12: Activation - ReLU
Key characteristics:
  • Five convolutional layers use same padding to maintain spatial dimensions
  • One convolutional layer uses valid padding to reduce dimensions before pooling
  • All convolutions use 3×3 kernels with stride of 1
  • ReLU activation functions introduce non-linearity after each convolution

​
Stage 2: Pooling and regularization

After the initial feature extraction, the model applies max pooling to reduce spatial dimensions and dropout for regularization. 
Layer 13: MaxPooling2D - pool size 2×2
Layer 14: Dropout - rate 0.25
The first dropout layer uses a rate of 0.25, randomly dropping 25% of neurons during training to prevent overfitting.

​
Stage 3: Deeper feature learning

The second convolutional block increases filter depth to 128, capturing more complex patterns. 
Layer 15: Conv2D - 128 filters, kernel 3×3, padding same
Layer 16: Activation - ReLU
Layer 17: Conv2D - 128 filters, kernel 3×3, padding valid
Layer 18: Activation - ReLU
Layer 19: MaxPooling2D - pool size 2×2
Layer 20: Dropout - rate 0.25
Design rationale:
  • Doubling the filter count to 128 allows the network to learn more abstract features
  • Second max pooling layer further reduces spatial dimensions
  • Dropout maintains regularization consistency

​
Stage 4: Classification head

The final layers flatten the feature maps and apply fully connected layers for classification. 
Layer 21: Flatten
Layer 22: Dense - 512 units, activation linear
Layer 23: Activation - ReLU
Layer 24: Dropout - rate 0.5
Layer 25: Dense - 7 units, activation linear
Layer 26: Activation - Softmax
Classification components:
  • Flatten layer: Converts 2D feature maps into 1D vector
  • Dense layer (512 units): Fully connected layer for high-level reasoning
  • Dropout (0.5): Aggressive regularization before final prediction
  • Output layer (7 units): One neuron per skin cancer category
  • Softmax activation: Converts outputs to probability distribution

​
Training configuration

​
Optimizer

  • Type: RMSprop
  • Learning rate: 0.0001
  • Decay: 1e-6
RMSprop (Root Mean Square Propagation) is particularly effective for CNNs as it adapts the learning rate for each parameter, helping the model converge more reliably.

​
Loss function

  • Type: Categorical crossentropy
  • Purpose: Measures the difference between predicted probability distributions and true labels

​
Initialization

  • Kernel initializer: Glorot Uniform (also called Xavier initialization)
  • Bias initializer: Zeros

​
Model capacity

​
Total parameters

The model contains millions of trainable parameters distributed across convolutional and dense layers:
  • Convolutional layers: Majority of parameters from 64-filter and 128-filter Conv2D layers
  • Dense layers: 512-unit fully connected layer contributes significant parameter count
  • Model size: Approximately 100MB across 25 binary shards

​
Architecture benefits

Deep feature hierarchy

Multiple convolutional layers learn increasingly abstract features, from edges and textures to complex lesion patterns.

Regularization

Dropout layers (0.25 and 0.5) and validation splits prevent overfitting on medical imagery.

Spatial efficiency

Max pooling reduces dimensions while preserving important features, making the model computationally efficient.

Web-ready

TensorFlow.js format enables browser-based inference without server dependencies.

​
Activation functions

​
ReLU (Rectified Linear Unit)

Used throughout the network for hidden layers: 
f(x) = max(0, x)
Advantages:
  • Computationally efficient
  • Reduces vanishing gradient problem
  • Introduces non-linearity for complex pattern learning

​
Softmax

Used in the final output layer: 
f(x_i) = exp(x_i) / Σ exp(x_j)
Purpose:
  • Converts raw scores to probabilities
  • Ensures output sums to 1.0
  • Enables multi-class classification interpretation

​
Architecture comparison

This architecture is inspired by VGG-16 but adapted for smaller dermatological images:
AspectVGG-16 StandardThis Model
Input size224×224×375×100×3
Convolutional blocks5 blocks2 blocks
Max pooling layers52
Dense layers32
Output classes1000 (ImageNet)7 (skin lesions)
DropoutMinimal0.25, 0.25, 0.5
The reduced input size and simplified architecture make this model more suitable for web deployment while maintaining strong classification performance on skin cancer images.

​
Next steps

Training process

Learn about the dataset and training methodology

Classifications

Understand the 7 skin cancer categories the model detects

Build docs developers (and LLMs) love

Get started for freeTalk to us