Comparative Evaluation of Deep and Lightweight CNN Architectures for Multiclass Classification of Rickettsia Skin Rashes

Abstract

Rickettsia diseases are frequently underdiagnosed due to nonspecific dermatological manifestations and the scarcity of publicly available annotated datasets. This study evaluates three pre-trained convolutional neural networks (CNNs)—ResNet-50, MobileNetV2, and AlexNet—for multiclass classification of Rickettsia skin rashes against six visually similar conditions: Chickenpox, Cowpox, Hand-Foot-and-Mouth Disease (HFMD), Healthy skin, Measles, and Monkeypox. A balanced dataset of 700 images (100 per class) was used, with five-fold cross-validation, data augmentation, and class weighting to mitigate data limitations and class imbalance. Results indicate that lightweight architectures, particularly MobileNetV2 and AlexNet, outperform deeper networks in detecting Rickettsia, achieving F1-scores of 93.14% and 93.88%, respectively, compared to 87.11% for ResNet-50. Ensemble prediction further enhanced stability and discrimination for rare classes. The findings suggest that computationally efficient CNN architectures provide a robust framework for early-stage screening of Rickettsia infections in low-resource clinical settings.