Datasets and Raw Results
Download
- DATASET (Figshare.com; 4.3 giga bytes)
- RAW RESULT (onedrive.com; 360 mega bytes)
- Virtual E Dataset (Figshare.com; 1.6 giga bytes)

1) DATASET
- Train dataset (thumanails)
A1 - Asan dataset (49567 images) - Clinical diagnosis (image findings +/- chart review)
A2 - Asan dataset (3741 images) - Clinical diagnosis (chart review)

- Validation dataset (1358 images)
B1 - Inje dataset (100 images) - Culture
B2 - Inje dataset (194 images) - KOH, or Culture, or Responsiveness to medications (TA ILI or antifungal drugs)
C - Hallym dataset (125 images) - KOH, or Culture, or Responsiveness to medications (TA ILI or antifungal drugs)
D - Seoul dataset (939 images) - KOH

2) RAW RESULT
- Human result
- AI result
- Test dataset and pdf
3) Virtual E Dataset
- Virtual dataset
E - Web dataset (3317 images) - Diagnosis predicted by CNNs (ResNet-152 + VGG-19; arithmatic mean of both outputs; training dataset: A1)

We created the E dataset to assess the semisupervised learning performance by conducting a Web-based image search for “tinea,” “onychomycosis,” “nail dystrophy,” “onycholysis,” and “melanonychia” in English, Korean, and Japanese on http://google.com and http://bing.com, and downloaded a total of 15,844 images.
From these images, the R-CNNs created a nail dataset of 3,317 images, since we had to discard many images because of low image resolution. The CNNs (model: ResNet-152 + VGG-19; arithmetic mean of both outputs; training dataset: A1) automatically classified images generated by the R-CNNs into six classes (760 onychomycosis, 1,316 nail dystrophy, 363 onycholysis, 185 melanonychia, 424 normal, and 269 others).