DeepBiteNet: A Lightweight Ensemble Framework for Multiclass Bug Bite Classification Using Image-Based Deep Learning.

Area of Science:

• Dermatology
• Artificial Intelligence
• Computer Vision

Background:

• Accurate insect bite identification from skin images is challenging due to subtle differences between bite types, varied skin reactions, and inconsistent image quality.
• Existing methods struggle with the nuances of diverse insect bites and image variations.

Purpose of the Study:

• To develop DeepBiteNet, a novel ensemble deep learning model for robust multiclass classification of insect bites from RGB images.
• To improve the accuracy and generalizability of automated insect bite identification.

Main Methods:

• An ensemble model, DeepBiteNet, was created by aggregating three diverse convolutional neural networks (DenseNet121, EfficientNet-B0, MobileNetV3-Small) using a stacked meta-classifier.
• A dataset of 1932 labeled insect bite images (eight classes) was used for training and evaluation.
• A domain-specific augmentation pipeline incorporated variations in lighting, occlusion, and skin tone to enhance model generalizability.

Main Results:

• DeepBiteNet achieved high accuracy: 89.7% (training), 85.1% (validation), and 84.6% (testing).
• The model outperformed fifteen benchmark CNN architectures in precision (0.880), recall (0.870), and F1-score (0.875).
• Optimized for mobile deployment using TensorFlow Lite, enabling efficient on-client computation.

Conclusions:

• Ensemble learning combined with realistic data augmentation significantly improves the reliability and usability of automated insect bite diagnosis.
• DeepBiteNet provides a strong foundation for mobile health (mHealth) solutions, aiding early diagnosis in underserved regions.