Dentomni
From a research question to dental image inference.
Dentomni investigates two ways to read a dental radiograph: classify the whole image, or locate individual lesions. The study compares standalone ResNet-50 with Faster R-CNN using a ResNet-50 feature pyramid backbone.
I built the Faster R-CNN detector; Rushil Thampan developed the standalone classifier. Our research report, training plots, prediction examples, and inference interface are available together in a public demo.
Knowing something is present is different from knowing where it is.
Dentomni studies that distinction in dental radiographs, comparing image classification with lesion localization under limited compute.
Two architectures
Standalone ResNet-50 predicts at image level. Faster R-CNN with ResNet-50-FPN adds bounding-box localization.
An inspectable experiment
Training curves, feature maps, prediction examples, and a downloadable comparative report accompany the demo.
Consumer hardware
The detector was trained for 70 epochs on an RTX 4060 laptop GPU with 8 GB of memory.
A flag, or a location?
A panoramic radiograph contains many teeth and regions. An image classifier compresses that information into an image-level prediction. A detector proposes regions and classifies them separately, adding coordinates that can be overlaid on the image. Dentomni explores the practical difference for caries and periapical lesions.
The comparative study is co-authored by Shivank Prabhudessai and Rushil Thampan at DBCE Goa. I developed the Faster R-CNN detector with a ResNet-50 feature pyramid backbone; Rushil developed the standalone ResNet-50 classifier. The deployed interface connects the experiment with model selection, image inference, and a public report library.
Small lesions make the details matter.
The detector starts from torchvision’s COCO-pretrained Faster R-CNN, with a prediction head adapted to the dental classes. Its feature pyramid exposes multiple spatial scales, while smaller anchors help address small target regions. Training uses a batch size of two and mixed precision to fit the laptop GPU.
Class imbalance is a central constraint: the original annotations contain substantially more caries examples than periapical lesions. The study introduces oversampling for the minority class and uses augmentations that preserve the relationship between the radiograph and its bounding boxes. The 70-epoch training history captures the staged training process.
The evaluation protocol belongs beside the score.
The report gives 87.9% detection accuracy after hybrid test-time augmentation and per-class non-maximum suppression, with mAP@0.5 of 0.2673. The detector evaluation uses 125 annotated images from a 200-image held-out set; 75 unannotated images are excluded. The final number therefore depends on the available annotations and the stated post-processing setup.
The standalone ResNet-50 report gives 80.0% image-level classification accuracy and macro F1 of 0.681 on 200 test images. Classification accuracy and detection accuracy are different measurements, and the evaluation sets differ. The table below presents the reported results with that context instead of treating their percentage difference as a controlled, like-for-like gain.
The mAP result is especially important because it exposes limits that a single headline accuracy can hide. Stronger evidence would require a larger, independently annotated evaluation set, consistent metric definitions, and prospective external validation. Dentomni is a research prototype; the report describes that validation as future work.
Reported results, with the conditions attached.
| Measurement | Faster R-CNN + ResNet-50-FPN | Standalone ResNet-50 |
|---|---|---|
| Task | Lesion detection + localization | Image-level classification |
| Reported accuracy | 87.9% detection accuracy | 80.0% classification accuracy |
| Additional metric | mAP@0.5: 0.2673 | Macro F1: 0.681 |
| Evaluation scope | 125 annotated / 200 held-out images | 200 test images |
| Output | Class predictions + bounding boxes | Image-level class prediction |
| Training hardware | RTX 4060 laptop · 8 GB | RTX 3060 desktop |
Source: the project’s comparative report, sections V–VIII. Detector accuracy includes hybrid TTA and per-class NMS. These are author-reported research results across different tasks and evaluation setups, not a clinical validation or a matched benchmark ranking.
How it comes together.
Prepare the data
Use annotated dental radiographs, preserve box geometry, and address class imbalance.
Train each approach
Learn image-level classification and multi-scale region detection as distinct tasks.
Evaluate & inspect
Read scores alongside predictions, loss curves, feature maps, and the test protocol.
Expose the research
Publish the report and inference interface so the experiment can be examined.
A closer look.
Explore the interface, the work, and the details behind it.
A result becomes useful when it can be examined.
Dentomni brings the model, its predictions, and its evaluation material into one public artifact. The key engineering contribution is spatial detection under constrained compute; the key research lesson is to keep the metric and its limitations together.