Table 4 Hyperparameters details.
From: Advancing skin cancer diagnosis with deep learning and attention mechanisms
Hyperparameter | Value | Description/range |
|---|---|---|
Learning Rate | 0.001 | Optimizer learning rate, chosen through grid search over [0.0001, 0.001, 0.01] |
Batch Size | 32 | Batch size used during training, selected from [16, 32, 64] based on memory constraints |
Epochs | 50 | Number of epochs, selected based on convergence time and model performance |
Optimizer | Adam | Adam optimizer was selected for its adaptive learning rate properties. We also experimented with RMSProp and SGD, but Adam outperformed them in terms of training stability and final accuracy. |
Dropout Rate | 0.5 | Dropout is used to regularize the model and prevent overfitting. Tested values in [0.3, 0.5, 0.7]. |
Weight Decay (L2 Reg.) | 0.0001 | L2 regularization to avoid overfitting, set after grid search across [0, 0.0001, 0.001] |
Activation Function | ReLU | The ReLU activation function, found to be effective for this segmentation task after comparisons with Leaky ReLU, Sigmoid, and Tanh. |
Loss Function | Dice Loss | Dice loss was used for its ability to handle imbalanced data and penalize incorrect segmentation of lesion boundaries. |
Input Image Size | 224 × 224 | Input image size, chosen based on available computational resources and the nature of the dataset. |
Number of Filters (Conv.) | 64 | Number of convolutional filters, chosen for a balance between model complexity and performance. |
Pooling Size | 2 × 2 | Standard 2 × 2 pooling was used to reduce spatial dimensions while retaining important features. |
