Computer-Aided Diagnosis System for Invasive Oral Cancer Detection using Deep Learning Techniques

This project presents a computer-aided diagnosis system for automated detection of invasive oral cancer from oral cavity images using deep learning. A transfer learning-based Convolutional Neural Network using MobileNetV2 is employed to classify images into Cancer and Normal categories. The system is designed for academic and research purposes to assist in early screening and decision support.

Model Architecture and Methodology

  • Base Model: MobileNetV2 (Transfer Learning)
  • Pre-trained on: ImageNet
  • Input image size: 224 x 224 x 3 (RGB)
  • Architecture: Global Average Pooling followed by Dense layers
  • Task: Binary classification (Cancer vs Normal)
  • Training: Optimized using Adam optimizer and Binary Cross-Entropy loss
  • Post-processing: Threshold tuning applied to reduce false negatives
  • Deployment: Final model deployed without retraining

Dataset and Model Performance

Dataset Sources:

Dataset Summary:

  • Multiple public datasets were merged to improve diversity
  • Total images used after merging datasets: 1988
  • Cancer images: 1185 | Normal images: 803
  • Dataset split into training, validation, and test sets

Performance Metrics (Test Set):

  • Accuracy: 95%
  • Precision (Cancer): 0.94
  • Recall (Cancer): 0.96
  • F1-score (Cancer): 0.95

Confusion Matrix Summary:

True Positives: 171 | False Negatives: 7 — False Positives: 10 | True Negatives: 141

Recall was prioritized to reduce false negatives due to the critical medical importance of missed cancer cases.

Image Upload Guidelines

  • Upload clear oral cavity images only
  • Supported formats: JPG, JPEG, PNG
  • Image must be RGB (3-channel)
  • Recommended resolution: minimum 224 x 224 pixels
  • Avoid blurred or low-light images
  • One image at a time

Upload Image

Choose image from your device

Upload an image above to see the preview and the Predict button below.

Sample Images

Try sample image from here

Disclaimer: This system is developed strictly for academic and research purposes. The predictions generated by this application do not constitute medical advice and must not be used as a substitute for professional clinical diagnosis.