Downstream Evaluation · Classification

DOVE Classification Performance

Evaluation across single-label and multi-label oral photography benchmarks.

01
Single-label classification

Few-shot Transfer

Protocol

Frozen DOVE encoder → image embeddings → L2 normalization → cosine 1-NN classifier

1-shot5-shot10-shot20-shot
Dataset 01

ORAL CANCER

Classes
2
Images
950

This dataset, comprises 500 oral cancer images and 450 non-cancer oral images.

results image
Figure 1 Four-model comparison on ORAL CANCER.
Dataset 02

Caries-Spectra

Classes
3
Images
2,000

The dataset comprises 2000 original JPG images across three categories: Advanced Enamel Caries (n=800), Early Stage Enamel Caries (n=800), and No Enamel Caries (n=400).

results image
Figure 2 Four-model comparison on Caries-Spectra.
\
Dataset 03

Oral Lesions: Malignancy Detection Dataset

Classes
2
Images
323

The dataset contains images of 165 benign lesions and 158 malignant lesions.

results image
Figure 3 Four-model comparison on Oral Leisions: Malignancy Detection Dataset.
Dataset 04

Oral Disease Sub

Classes
3
Images
1,799

The dataset contains 6,613 images of calculus, hypodontia and ulcer

results image
Figure 4 Four-model comparison on Oral Disease Sub.
Dataset 05

Dataset-of-oral-mucosal-diseases

Classes
5
Images
1348

The dataset comprises 1,348 intraoral mucosal images categorized into five classes: normal oral mucosa, oral leukoplakia, oral lichen planus, oral submucous fibrosis, and oral cancer.

results image
Figure 5 Four-model comparison on Dataset-of-oral-mucosal-diseases.
02
SINGLE-LABEL CLASSIFICAtiON

End-to-End Finetuning

Protocol

Trainable encoder + classification head→ Supervised fine-tuning

5%10%20%30%50%100%
Dataset 01

ORAL CANCER

Classes
2
Images
950

This dataset, comprises 500 oral cancer images and 450 non-cancer oral images.

results image
Figure 1 Four-model comparison on ORAL CANCER.
Dataset 02

Caries-Spectra

Classes
3
Images
2,000

The dataset comprises 2000 original JPG images across three categories: Advanced Enamel Caries (n=800), Early Stage Enamel Caries (n=800), and No Enamel Caries (n=400).

results image
Figure 2 Four-model comparison on Caries-Spectra.
\
Dataset 03

Oral Lesions: Malignancy Detection Dataset

Classes
2
Images
323

The dataset contains images of 165 benign lesions and 158 malignant lesions.

results image
Figure 3 Four-model comparison on Oral Leisions: Malignancy Detection Dataset.
Dataset 04

Oral Disease Sub

Classes
3
Images
1,799

The dataset contains 6,613 images of calculus, hypodontia and ulcer

results image
Figure 4 Four-model comparison on Oral Disease Sub.
Dataset 05

Dataset-of-oral-mucosal-diseases

Classes
5
Images
1348

The dataset comprises 1,348 intraoral mucosal images categorized into five classes: normal oral mucosa, oral leukoplakia, oral lichen planus, oral submucous fibrosis, and oral cancer.

results image
Figure 5 Four-model comparison on Dataset-of-oral-mucosal-diseases.