LUNG CANCER CLASSIFICATION USING VGG16

VGG16 Model Overview:

Architecture:

• Layers: 16 weight layers (13 convolutional layers + 3 fully connected layers).
• Conv Layers: Small 3x3 receptive fields, stacked for deeper architecture.
• Pooling: Max-pooling layers (2x2) follow some of the conv layers.
• FC Layers: Three fully connected layers at the end, with softmax activation in the final layer for classification.

Design Principles:

• Use of small convolutional filters (3x3) and consistent design.
• Trained on ImageNet in 2014.

Considerations:

• Can only work with RGB images
• Expects image size of 224 x 224 pixels.

TECHNICAL IMPLEMENTATION

1. Data Preparation:

• Loaded a Lung Cancer Classification dataset from Kaggle.
• Resized images to 224x224 pixels to match the input size expected by VGG16.
• Kept the channel dimension to 3 as expected by VGG16 model
• Used os module for path construction

2.Model Architecture:

• Used VGG16 pre-trained on ImageNet as the feature extractor.
• Added custom top layers including Global Average Pooling and Dense layers to adapt the model for our specific classification task.

3. Transfer Learning:

• Frozen the pre-trained VGG16 layers to retain learned features.
• Fine-tuned the custom top layers to adapt to the new dataset.

4. Training:

• Split the dataset into training and testing sets (80/20 split).
• Trained the model using categorical cross-entropy loss and Adam optimizer.
• Evaluated the model on the test set, achieving accuracy of 98%.