GingivitisSeg: A Public Dataset and Comparative Benchmark for Gingivitis Segmentation in Intraoral Images
Organizations: Department of Artificial Intelligence and Data Engineering, Ankara University, Ankara, Türkiye · Department of Periodontology, Faculty of Dentistry, Ankara University, Ankara, Türkiye · Department of Computer Engineering, Ankara University, Ankara, Türkiye · Department of Periodontology, Faculty of Dentistry, Lokman Hekim University, Ankara, Türkiye
Abstract
Gingivitis segmentation from intraoral photographs can support computer-assisted periodontal assessment, but pixel-level delineation remains challenging due to severe foreground-background class imbalance, low-contrast boundaries, and variations in illumination and anatomy. We introduce GingivitisSeg, a public benchmark comprising 961 intraoral images with pixel-level annotations of gingival inflammation and predefined training (672), validation (144), and test (145) partitions. Seven semantic segmentation model-backbone configurations were compared using standardized partitions, input resolution, and evaluation metrics. Controlled ablation experiments examined class-weighted cross-entropy loss, Dice loss, and auxiliary supervision. UPerNet with ResNet50 achieved the highest mean Intersection over Union (mIoU) of 76.49% and gingivitis IoU of 54.91%. However, differences from DeepLabV3+ with ResNet50 and SegFormer with MiT-B2 were not statistically significant after Holm correction. The evaluated optimization components improved performance when combined in the UPerNet-ResNet50 configuration. Further analyses showed that performance was associated with gingivitis region size and decreased under stronger image blur. The dataset, benchmark splits, and source code are available at https://github.com/zeynepozdemir/GingivitisSeg.
Figures & tables
| Study | Image Type | Task | Size | Annotation | Focus | Evaluation |
|---|---|---|---|---|---|---|
| Li et al. (2021) [ 4 ] | Intraoral Photo | Gingivitis Seg. | 337 + 110 | Pixel-level | Disease Localization | Single Model |
| Chau et al. (2023) [ 5 ] | Frontal Photo | Gingivitis Detection | 567 | Pixel-level | Clinical Diagnosis | Single Model |
| Cheng et al. (2024) [ 6 ] | Intraoral Photo | Mobile Gingivitis Detection | 415 | Pixel-level | Mobile Deployment | Deployment |
| Chau et al. (2025) [ 7 ] | Smartphone Photo | Clinical Detection | 44 Patients | Clinical Assessment | External Validation | Clinical Validation |
| Cheng et al. (2025) [ 8 ] | Mixed Oral Images | Segmentation + Filtering | 332 + 83 | Pixel-level | VLM Assisted Workflow | Single Model |
| Liu et al. (2024) [ 2 ] | Endoscopic Oral Image | Multi disease Seg. | 3365 | Pixel-level | Oral Disease Analysis | U-Net Comparison |
| Property | Value |
|---|---|
| Number of images | 961 |
| Source composition | 432 clinical + 529 public-source images |
| Image type | Intraoral photographs |
| Classes | Background, Gingivitis Area |
| Annotation type | Pixel-level masks |
| Native image width | 416–7286 pixels |
| Model | Backbone | mIoU | mAcc | mDice | aAcc |
|---|---|---|---|---|---|
| FCN | U-Net [ 11 ] | 73.17 1.27 | 86.67 4.00 | 82.13 1.06 | 97.77 0.45 |
| DeepLabV3 [ 12 ] | ResNet50 [ 18 ] | 75.49 0.46 | 91.12 0.12 | 84.15 0.37 | 97.92 0.08 |
| DeepLabV3+ [ 13 ] | ResNet50 [ 18 ] | 76.25 0.24 | 90.65 0.59 | 84.77 0.20 | 98.06 0.01 |
| SegFormer [ 14 ] | MiT-B0 | 73.95 1.03 | 92.35 1.48 | 82.87 0.86 | 97.59 0.27 |
| SegFormer [ 14 ] | MiT-B2 | 75.74 0.51 | 90.96 1.03 | 84.35 0.41 | 97.96 0.13 |
| UPerNet [ 15 ] | ResNet50 [ 18 ] | 76.49 0.40 | 90.33 0.04 | 84.96 0.32 | 98.11 0.06 |
| Model | Backbone | Background | Gingivitis | ||||
|---|---|---|---|---|---|---|---|
| IoU back | Acc back | Dice back | IoU gin | Acc gin | Dice gin | ||
| FCN | U-Net [ 11 ] | 97.72 0.46 | 98.42 0.71 | 98.85 0.23 | 48.63 2.09 | 74.92 8.70 | 65.42 1.88 |
| DeepLabV3 [ 12 ] | ResNet50 [ 18 ] | 97.87 0.08 | 98.33 0.09 | 98.92 0.04 | 53.11 0.84 | 83.90 0.33 | 69.37 0.71 |
| DeepLabV3+ [ 13 ] | ResNet50 [ 18 ] | 98.02 0.01 | 98.50 0.03 | 99.00 0.01 | 54.50 0.47 | 82.78 1.20 | 70.55 0.40 |
| SegFormer [ 14 ] | MiT-B0 | 97.53 0.27 | 97.90 0.37 | 98.75 0.14 | 50.37 1.79 | 86.80 3.32 | 66.99 1.58 |
| SegFormer [ 14 ] | MiT-B2 | 97.92 0.13 | 98.38 0.20 | 98.95 0.06 | 53.57 0.88 | 83.54 2.25 | 69.76 0.75 |
| Comparator | mIoU | mDice | IoU gin | Dice gin |
|---|---|---|---|---|
| FCN/U-Net | ||||
| DeepLabV3 with ResNet50 | 0.004 | 0.007 | 0.005 | 0.009 |
| DeepLabV3+ with ResNet50 | 0.730 | 0.797 | 0.637 | 0.741 |
| SegFormer with MiT-B0 | ||||
| SegFormer with MiT-B2 | 0.0518 | 0.0503 | 0.0576 | 0.0567 |
| UPerNet with ConvNeXt-Tiny | 0.009 | 0.024 | 0.031 | 0.042 |
| Study | Task | Model | IoU / mIoU | Accuracy |
|---|---|---|---|---|
| Li et al. (2021) [ 4 ] | Gingival disease segmentation | DeepLabV3+ | 0.650 a / 0.379–0.485 b | – |
| Chau et al. (2023) [ 5 ] | Gingivitis detection | Deep learning segmentation model | 0.600 | – |
| Cheng et al. (2024) [ 6 ] | Mobile gingivitis detection | Mobile deep learning model | – | 0.900 |
| Chau et al. (2025) [ 7 ] | External validation | Deep learning gingivitis detection model | – | 0.850 |
| Cheng et al. (2025) [ 8 ] | Gingival disease segmentation | GumAgent | 0.482 | 0.886 c |
| Liu et al. (2024) [ 2 ] | Segmentation of multiple oral diseases | Oral-Mamba | 0.700 d | 0.830 d |
| Optimization Components | Overall Performance | Gingivitis Class | |||||
|---|---|---|---|---|---|---|---|
| Configuration | Class Wt. | Dice Loss | Aux. Sup. | mIoU | mDice | IoU gin | Dice gin |
| Full | 76.77 | 85.18 | 55.43 | 71.33 | |||
| w/o Class Wt. | 75.90 | 84.41 | 53.45 | 69.66 | |||
| w/o Dice Loss | 73.63 | 82.61 | 49.87 | 66.55 | |||
| w/o Aux. Sup. | 75.31 | 84.01 | 52.83 | 69.14 | |||
| CE Only | 75.15 | 83.79 | 52.05 | 68.46 | |||
| Foreground Group | Mean Area (%) | IoU gin (%) | Dice gin (%) | |
|---|---|---|---|---|
| Small | 48 | 1.13 | 31.30 19.01 | 44.47 22.92 |
| Medium | 48 | 2.66 | 47.00 18.50 | 61.43 20.67 |
| Large | 49 | 6.79 | 59.38 11.85 | 73.74 10.74 |
| Preprocessing | mIoU | mAcc | mDice | aAcc | IoU gin | Acc gin | Dice gin |
|---|---|---|---|---|---|---|---|
| Direct resize ( ) | 76.77 | 90.36 | 85.18 | 98.15 | 55.43 | 82.11 | 71.33 |
| Aspect ratio preservation with padding | 74.41 | 89.84 | 83.23 | 97.81 | 51.06 | 81.41 | 67.60 |
| Gingivitis IoU | Gingivitis Dice | |||||
|---|---|---|---|---|---|---|
| Condition | Mean | Mean | ||||
| Clean | 45.99 | – | – | 59.98 | – | – |
| Gaussian ( ) | 45.09 | 0.646 | 58.81 | 0.699 | ||
| Gaussian ( ) | 41.94 | 0.025 | 54.94 | 0.030 | ||
| Gaussian ( ) | 38.60 | 0.0011 | 50.74 | 0.00085 | ||
| Motion ( ) | 44.25 | 0.646 | 57.55 | 0.699 | ||
Medical Image Segmentation
Medical image segmentation delineates organs, lesions, and other structures in scans to support measurement, diagnosis, treatment planning, and research.
See all 839 papers