Hybrid Medical Expert System for Skin Disease Classification

The Hybrid Medical Expert System is a skin disease classification system integrating deep learning models with rule-based expert systems. Built using Streamlit, the system provides an intuitive interface for uploading skin lesion images, generating predictions, and giving domain-specific recommendations.

---

Key Features

1. Hybrid Approach: Combines deep learning with rule-based reasoning for enhanced interpretability.
2. Multi-Model Implementation: Utilizes state-of-the-art models including MobileNet, DenseNet121, and EfficientNetB3.
3. Web Interface: Easy-to-use application for real-time image analysis, visualization, and prediction explanation.

---

Datasets

1. Kaggle Skin Disease Dataset

• Description: Comprises 900 images across 9 disease categories.
• Preprocessing:
  • Images resized to 256x256.
  • Data augmentation applied to improve generalization.
• Train-Test Split: 80:20.

2. ISIC Skin Disease Dataset

• Description: Contains 25,331 images across 8 disease categories.
• Preprocessing:
  • Handled significant class imbalance using oversampling and undersampling, resulting in ~4,000 samples per class.
  • Images resized to 256x256.
  • Augmentation applied.
• Split Ratio: 70:15:15 for training, validation, and testing.

---

Model Implementations and Results

1. MobileNet

• Purpose: Lightweight model for mobile and resource-constrained environments.
• Dataset: Kaggle Skin Disease Dataset.
• Results:
  • Train Accuracy: 99.18%
  • Validation Accuracy: 7.14% (indicates overfitting)
  • Train Loss: 0.0896
  • Validation Loss: 4.0363

---

2. DenseNet121

• Purpose: Dense connections for efficient gradient flow.
• Dataset: Kaggle Skin Disease Dataset.
• Results:
  • Train Accuracy: 76.30%
  • Validation Accuracy: 31.43% (indicates overfitting)
  • Test Accuracy: 49.17%
  • Train Loss: 0.3048
  • Validation Loss: 0.5789

---

3. EfficientNetB3

• Purpose: Scalable and accurate model for large datasets.
• Dataset: ISIC Skin Disease Dataset.
• Results:
  • Train Accuracy: 96.09%
  • Test Accuracy: 82.43%
  • Precision: 82.43%
  • Recall: 82.43%
  • F1 Score: 0.8242

---

System Architecture

1. Deep Learning Module:
   • Predicts disease category using pre-trained models fine-tuned for medical images.
2. Knowledge-Based Module:
   • Uses predefined rules from domain knowledge to provide recommendations and explanations.
3. Streamlit Interface:
   • Facilitates user interaction with features for image upload, prediction display, and interpretability tools like saliency maps.

---

Demo Screenshots

Image Upload and Prediction

Prediction Results with Recommendations

---

Challenges and Solutions

1. Data Imbalance:

   • Resolved using oversampling and undersampling techniques.

2. Model Overfitting:

   • Tackled through augmentation and rigorous hyperparameter tuning.

3. Computational Intensity:

   • Optimized integration of deep learning and rule-based systems for faster predictions.

---

Future Work

1. Expand Knowledge Base:
   • Improve the rule-based system with a broader range of conditions.
2. Edge Deployment:
   • Optimize lightweight models for deployment on mobile devices.

---

Citation

Developed by: Sharif Ehab

Institution: Cairo University, Faculty of Engineering (Class 2025)