Viral Rash Classification

Introduction

This project aims to classify three infectious skin diseases - Chickenpox, Measles, and Monkeypox - using various deep learning models. The goal is to compare the performance of different architectures on a relatively small dataset of medical images.

Dataset

• Total classes: 3 (Chickenpox, Measles, Monkeypox)
• Training data: 160 images per class (480 total)
• Validation data: 33 images per class (99 total)

Analysis

Loss Curve

Accuracy Curve

Confusion Matrix

1. Custom-CNN

   • Best Validation Accuracy: 57.58%

   • Validation Loss: 1.5858

2. Inception-ResNet-V2

   • Best Validation Accuracy: 92.93%

   • Validation Loss: 0.2921

3. ResNet-50

   • Best Validation Accuracy: 91.92%

   • Validation Loss: 0.3740

4. NesNet-A-Large

   • Best Validation Accuracy: 85.86%

   • Validation Loss: 0.6669

YOLO11 Classification

Confusion Matrix

Accuracy Curve/Loss Curve

5. YOLO11N Classification

   • Best Validation Accuracy: 82.8%

   • Validation Loss: 0.7135

Conclusion

Based on the comparative analysis between Inception-ResNet-V2 with 93.94% accuracy and ResNet-50 models with 91.92% accuracy, ResNet-50 emerges as the more practical choice despite its slightly lower peak validation accuracy. ResNet-50 demonstrates superior efficiency with faster training times (~1-1.5 minutes per epoch vs 2-3 minutes), better generalization with less overfitting, and more stable validation metrics.

While Inception-ResNet-V2 achieves marginally higher peak accuracy, it suffers from more pronounced overfitting after Epoch 6 and shows unstable validation loss patterns. ResNet-50's simpler architecture requires fewer computational resources while maintaining competitive performance, making it the recommended model for most applications where the 2% accuracy difference isn't critical. Its combination of efficiency, stability, and strong performance makes it a more practical choice for real-world applications.