Ensemble Deep Learning Approach for Nail Disease Classification Using EfficientNet-V2 and ConvNeXt-V2

Abstract: A broad range of human nail disorders affects the health, texture, and appearance of toenails and fingernails. Various reasons such as injury, infection, systemic diseases, and nutrition deficiency can lead to these nail disorder conditions. Additionally, the nails might be impacted by certain skin dis-orders like psoriasis and eczema. Since nails are associated with good grooming habits and hygiene, the quality of nails plays an important role in creating an initial impression. Hence, providing a timely and accurate diagnosis provides satisfactory results for the patient. Many traditional diagnosis methods of nail disorders, such as nail scraping or clipping for lab experiments, dermatoscopy, and clinical observations, are very time-consuming, often subjective to bias, involve many steps, and require the assistance of a specialist. In recent years, technology including digital photography and computer vision have opened up numerous opportunities to increase efficiency by way of automation in this area through the use of digitization and Deep Learning algorithms, enabling more objective and timely assessments of nails in regard to abnormalities. Research in deep learning has shown positive performance in medical image classification by using Convolutional Neural Networks (CNNs). This study proposes an ensemble technique using EfficientNet-V2 and ConvNeXt-V2 as the CNN architectures for enhancing the performance of nail disease classification. ConvNeXt-V2 scored 98.51% on the test, compared to 95.12% for EfficientNet-V2. The suggested fusion ensemble performs noticeably better than individual backbone models, according to experimental findings. The feature-level fusion ensemble proposed in this study further enhanced the results, giving a test accuracy of 99.04% with a heightened degree of robustness. By studying the performance between these two state-of-the-art architectures, we expect to see improvements in both the performance and generalization of performance in actual clinical settings.