Introduction

In this blog post, I share a complete overview of a deep learning-based medical imaging project focused on detecting leprosy from skin images. We use a blend of advanced CNN architectures like EfficientNet, VGG16, MobileNet, and Xception, all optimized through hinge loss—a method inspired by SVMs—to accurately distinguish between leprosy and non-leprosy images.

Pipeline Architecture

🔗 Project Links

Dataset Overview

The dataset is structured into two classes:

  • Leprosy (Label 1)
    Images show visible symptoms of leprosy.
    Folder: new_train/Leprosy

  • Non-Leprosy (Label 0)
    Images depict healthy skin.
    Folder: new_train/Non_Leprosy

Data Preprocessing

We applied several preprocessing steps to ensure consistent input and enhance generalization:

  • Resizing all images to (256, 256) pixels.
  • RGB Conversion using Image.open(...).convert("RGB").
  • Random Saturation variation using HSV channel manipulation.
  • Clean Directory Handling with os.path.join() and directory checks.
  • Systematic File Renaming to avoid duplicates and maintain order.

Modeling Approach

We tested several CNN architectures using transfer learning:

  • EfficientNetB0
    Pre-trained model fine-tuned for our binary classification task.

  • VGG16
    Simple and widely used; enhanced with hinge loss and linear output.

  • MobileNet
    Lightweight and ideal for embedded/mobile applications.

  • Xception
    Leverages depthwise separable convolutions for efficiency.

All models pass through a final Dense layer (linear activation) with L2 regularization and use hinge loss to enforce margin-based classification—akin to SVMs.

If final output > 0 → Leprosy
If final output < 0 → Non-Leprosy

Why These Models?

  • Transfer learning leverages pre-trained features on ImageNet.
  • EfficientNet adapts quickly with minimal training data.
  • Hinge loss maximizes separation margin between classes.
  • L2 regularization prevents overfitting.

Training Strategy

  • Train/Test Split: 80% training, 20% testing
  • Optimizer: Adam with learning rate 0.01
  • Loss: Hinge loss
  • Regularization: L2

Evaluation Metrics

Metric Value
Precision 0.931
Recall 0.947
F1 Score 0.939

Confusion Matrix

Hosting the Model in Google Colab

  1. Upload the files:

    • modelapp.py
    • svm_in_cnn.h5
    • run.sh

  2. Run the command in your Colab notebook.

  3. Click the provided URL.

  4. Use your IP address as the password.

App Screenshot 1

App Screenshot 2

Final Thoughts

This project combines the strengths of traditional SVM theory and modern deep learning architectures. It’s an ideal example of how even small datasets can yield accurate results with the right preprocessing, model selection, and evaluation pipeline.

👨‍⚕️ Bridging the gap between machine learning and healthcare—one pixel at a time.