Skin Disease Detection System Using Deep Learning

Final Year Project
Author: Muhammad Haikhal Bin Omanudin Baki

Overview

This project implements a deep learning-based skin disease detection system capable of classifying 7 types of skin lesions using the HAM10000 dataset. The system features a custom CNN architecture with hair removal preprocessing and a user-friendly web interface for real-time predictions.

Objectives

1. Design and develop a deep learning-based model capable of detecting and classifying common skin diseases from image data.
2. Evaluate the performance and accuracy of deep learning-based models for skin disease detection.
3. Implement a web-based skin disease detection interface that enables users to upload skin images and view real-time prediction results.

Detectable Skin Conditions

Condition	Code	Description
Melanocytic nevi	nv	Common moles (benign)
Melanoma	mel	Serious skin cancer
Benign keratosis-like lesions	bkl	Seborrheic keratoses, solar lentigo
Basal cell carcinoma	bcc	Common skin cancer
Actinic keratoses	akiec	Pre-cancerous lesions
Vascular lesions	vasc	Angiomas, hemorrhage
Dermatofibroma	df	Benign skin growths

Project Structure

Project/
├── HAM10000_images_part_1/    # Dataset images (part 1)
├── HAM10000_images_part_2/    # Dataset images (part 2)
├── HAM10000_metadata.csv      # Dataset metadata
├── models/                    # Trained model files
│   ├── skin_disease_model.h5
│   ├── best_model.h5
│   └── class_indices.json
├── results/                   # Training results and visualizations
│   ├── training_history.png
│   ├── confusion_matrix.png
│   ├── dataset_distribution.png
│   └── classification_report.txt
├── train.py                   # Training script
├── preprocessing.py           # Image preprocessing module
├── app.py                     # Streamlit web application
├── requirements.txt           # Python dependencies
└── README.md                  # This file

Installation

Prerequisites

• Python 3.8 or higher
• pip package manager

Setup

1. Clone or navigate to the project directory:

   cd Project

2. Create a virtual environment (recommended):

   python -m venv venv
   
   # Windows
   venv\Scripts\activate
   
   # Linux/Mac
   source venv/bin/activate

3. Install dependencies:

   pip install -r requirements.txt

Usage

1. Training the Model

To train the CNN model on the HAM10000 dataset:

python train.py

This will:

• Load and preprocess the dataset
• Apply data augmentation and oversampling
• Train the CNN model with early stopping
• Generate evaluation metrics and visualizations
• Save the trained model to models/

Training Parameters (configurable in train.py):

• IMG_SIZE: 128x128 pixels
• BATCH_SIZE: 32
• EPOCHS: 25 (with early stopping)
• LEARNING_RATE: 0.0001

2. Running the Web Interface

After training, launch the Streamlit web application:

streamlit run app.py

The application will open in your browser at http://localhost:8501

Features:

• Upload skin lesion images (JPG, PNG)
• Real-time prediction with confidence scores
• Hair removal preprocessing visualization
• Disease information and recommendations
• All class probabilities display

Methodology

Data Preprocessing

1. Hair Removal (Inpainting):

   • Convert to grayscale
   • Apply morphological black-hat transform (17x17 kernel)
   • Create binary mask via thresholding
   • Inpaint using Telea algorithm

2. Data Augmentation:

   • Rotation (±20°)
   • Width/height shift (20%)
   • Shear transformation (20%)
   • Zoom (20%)
   • Horizontal flip

3. Class Balancing:

   • Oversampling minority classes to match majority class count

CNN Architecture

Input (128x128x3)
    ↓
Block 1: Conv2D(32) → Conv2D(32) → BatchNorm → MaxPool → Dropout(0.25)
    ↓
Block 2: Conv2D(64) → Conv2D(64) → BatchNorm → MaxPool → Dropout(0.25)
    ↓
Block 3: Conv2D(128) → Conv2D(128) → BatchNorm → MaxPool → Dropout(0.25)
    ↓
Block 4: Conv2D(256) → Conv2D(256) → BatchNorm → MaxPool → Dropout(0.25)
    ↓
Flatten → Dense(512) → BatchNorm → Dropout(0.5) → Dense(256) → Dropout(0.5)
    ↓
Output: Dense(7, softmax)

Training Callbacks

• Early Stopping: Monitor validation loss, patience=5
• ReduceLROnPlateau: Reduce learning rate by 0.5 after 3 epochs of no improvement
• ModelCheckpoint: Save best model based on validation accuracy

Evaluation Metrics

The model is evaluated using:

• Accuracy: Overall classification accuracy
• Confusion Matrix: Visualize prediction patterns
• Classification Report: Per-class precision, recall, F1-score

Results are saved in the results/ directory.

API Reference

preprocessing.py

# Apply hair removal to an image
from preprocessing import hair_removal
cleaned_image = hair_removal(image_path_or_array)

# Preprocess for model prediction
from preprocessing import preprocess_for_prediction
processed = preprocess_for_prediction(image, target_size=(128, 128))

Disease Information

from preprocessing import DISEASE_INFO
info = DISEASE_INFO['Melanoma']
print(info['description'])
print(info['severity'])
print(info['recommendation'])

Troubleshooting

Model not found error

Ensure you have trained the model first:

python train.py

Out of memory during training

Reduce batch size in train.py:

BATCH_SIZE = 16  # or 8

Slow training on CPU

Consider using Google Colab with GPU or reduce image size:

IMG_SIZE = 64

Disclaimer

⚠️ Important: This system is designed for educational and screening purposes only. It should NOT be used as a substitute for professional medical diagnosis. Please consult a qualified dermatologist for accurate diagnosis and treatment.

Dataset

This project uses the HAM10000 dataset (Human Against Machine with 10000 training images):

• 10,015 dermatoscopic images
• 7 diagnostic categories
• Collected over 20 years at various clinics

Citation:

Tschandl, P., Rosendahl, C. & Kittler, H. The HAM10000 dataset, a large collection of 
multi-source dermatoscopic images of common pigmented skin lesions. Sci Data 5, 180161 (2018).

License

This project is developed as part of a Final Year Project for educational purposes.

Contact

For questions or feedback, please contact:

• Name: Muhammad Haikhal Bin Omanudin Baki
• Project: Final Year Project (FYP)