Deep Learning-Based Automated Classification of Pathogenic Fungi from Microscopic Images using Enhanced EfficientNet-B4

This project presents a federated learning framework with attention‑based aggregation for microscopic fungal image classification. The system identifies five fungal species: Candida albicans, Aspergillus niger, Trichophyton rubrum, Trichophyton mentagrophytes, and Epidermophyton floccosum using the DeFungi dataset containing 9,114 images with class imbalance. Two deep learning architectures are implemented and compared: an EfficientNet‑B4 baseline and a Vision Transformer (ViT) combined with a Learnable Spatial‑Channel Attention (LSCA) module. The EfficientNet model achieves 91.52% test accuracy, while the ViT‑LSCA model attains 84.28% accuracy with improved per‑class robustness. Federated learning enables multiple clients to collaboratively train the model without sharing raw image data, preserving data privacy. An attention‑based aggregation mechanism optimally weights client updates. The project also includes a Flask web application with modules for user registration, login, fungal classification prediction, and logout, using HTML, CSS, and JavaScript for the frontend. Experimental results demonstrate that attention mechanisms help address class imbalance and that federated learning is suitable for privacy‑sensitive medical image classification tasks.

Keywords: Federated Learning, Attention Mechanism, Vision Transformer, LSCA, Fungal Classification, Microscopic Images, DeFungi Dataset, Privacy Preservation, Medical Image Analysis, Deep Learning