Objective

A deep learning-based project that classifies skin diseases using a custom Convolutional Neural Network (CNN). This project explores how well a CNN can learn from image data alone—leveraging data augmentation, multiple convolutional blocks, and extensive evaluation.The main aim of this project was to build and evaluate a custom CNN model capable of accurately classifying images of 20 different skin diseases using the "20 Skin Diseases Dataset" and analyze how CNNs perform on complex medical image data.

Dataset Used: 20 Skin Diseases Dataset

Structure:

• Images are organized into folders per class.
• The dataset contains over 10,000+ high-resolution skin images.
• Each image is labeled by disease type.

Project Overview

1. Data Loading & Preprocessing:

• Images loaded and resized to 192x192.
• Labels created using one-hot encoding.
• Data split into training and validation sets.
• Visualized random image-label pairs for verification.

2. Data Augmentation:

• Rotation, flipping, zoom, shear, and shifts applied using ImageDataGenerator.
• Normalization (rescaling) applied for better training convergence.

3. Model Architecture:

• Custom CNN with 4 convolutional blocks.
• Final layers include flattening, fully connected layers, and softmax output.
• Regularization through max-pooling and ReLU activation.

4. Training:

• Trained for 20 epochs using Adamax optimizer and categorical crossentropy.
• Used ModelCheckpoint to save model after each epoch.

5. Evaluation:

• Plotted training/validation accuracy and loss.
• Visualized predictions on test images.
• Generated and visualized a confusion matrix.

Dependencies & Libraries Used

• numpy : Array operations
• pandas : Data manipulation (minor use)
• matplotlib, seaborn: Data and result visualization
• opencv-python (cv2): Image loading and processing
• tensorflow: Building and training the CNN
• scikit-learn: Model evaluation (train_test_split, confusion_matrix)

Install libraries using:

 'pip install numpy pandas matplotlib seaborn opencv-python tensorflow scikit-learn' 

Model Performance:

• Final Training Accuracy: 95.5%
• Final Validation Accuracy: 96.9% (subject to variation per run)
• Loss Trends: Converged smoothly with minimal overfitting (1.2%)
• Confusion Matrix: Reveals confusion between similar-looking diseases (visualized below)

Observation and Conclusion:

CNNs are powerful in extracting spatial hierarchies in medical images. Data augmentation significantly boosts generalization in limited-data scenarios. Deeper convolutional stacks (like 4 blocks here) help the model grasp complex patterns, especially in high-res images.Even a custom CNN, without transfer learning, can perform well with proper tuning and preprocessing.

Results

Training Accuracy & Loss

Validation Accuracy & Loss

Predicted Images

Confusion Matrix

Precautions: Medical datasets often have class imbalance and inter-class similarity, which challenges even robust CNNs.