Table 1 Summary of related work for nail disease detection .
Reference | Method | Accuracy (%) | Dataset Used |
|---|---|---|---|
Ensemble of three CNNs for feature extraction and a Random Forest classifier | 84.58% | A custom dataset with 4,190 images | |
YOLOv5, Mask R-CNN, Basic CNN | 98.5% | Custom dataset with 3,018 images | |
Trained CNN model. | 95% | Custom dataset with 185 images | |
Ensemble model combining ResNet-152 and VGG-19 | area under the curve (AUC) of up to 0.98, with sensitivity and specificity results of 96.0% and 94.7% | Custom dataset of 49,567 images | |
U-NET-based segmentation model with ResNet-50 | sensitivity of 94% and a specificity of 77%, overall accuracy of 86.49% | 664 whole-slide images (WSIs) from four different laboratories | |
Deep Hybrid Learning (DHL) approach, combining a pre-trained DenseNet201 convolutional neural network for feature extraction and a traditional SGDClassifier for classification. | 94% | 821 images of nails | |
CNN-based model | 98.44%. | 340 nail images, representing eight different nail diseases. After augmentation, the dataset was expanded to 2720 images​ | |
transfer learning with pre-trained CNN models, specifically focusing on ResNet101V2 | 89% | It used 234 images (187 training images and 47 test images) of Kaggle that were resized to 128 × 128 pixels and augmented to 1309 training images all in total. | |
Convolutional Neural Networks (CNN) | 88.33%. | nail images in PNG format | |
K-Nearest Neighbors (KNN) classifier | 93.63% | 483 nail image samples | |
Transfer Learning with CNNs (AlexNet, Vgg16, GoogleNet, ResNet50, and DenseNet201) | ResNet50 & DenseNet201: 96.39% GoogleNet: 93.98% AlexNet: 92.5% Vgg16: 87.5% | 280 images, | |
hybrid CNN-LSTM model | 94% | 8,536 images | |
VGG16 model with transfer learning, adding two Artificial Neural Network (ANN) layers for classification | 94% | custom dataset consisting of 600 healthy nail images, 192 melanoma images, and 248 yellow nail syndrome images. | |
CNNs for feature extraction and classification. The model architecture was fine-tuned using AlexNet. | 87.33% | custom dataset consists of 18,025 images | |
CNNs and proposed a fusion model combining VGG16 and GoogleNet architectures using feature and decision-level fusions. | 97.50% with VGG16, while the fusion model reached a maximum accuracy of 89.48%. | 8033 nail images across eight classes after augmentation, sourced from Roboflow. | |
CNN with the Inception-v3 architecture, leveraging transfer learning for classification. | 95.24%. | 115 images of Terry’s nails and 100 images of healthy nails |
