Adravision (Adra - AI Dental Radiograph Assist, YC S21) builds dental diagnostic software that reads X-ray radiographs to help clinicians diagnose faster and more reliably, with charting, patient education, and clinic analytics built in.

Founded by scientists, dental practitioners, and AI experts, the company has earned FDA clearance for its Adravision Perio device. Its users are dental practices and insurance companies - audiences where diagnostic accuracy is non-negotiable.

Dental radiograph interpretation is a fine-grained detection problem: Adravision's models must localize many distinct features on a single X-ray - cavities, periapical infection, calculus, implants, and fillings - and the outputs feed directly into clinical and insurance decisions.

That puts a hard premium on accuracy, and shaped several constraints:

• Quality matters more than raw speed, since dentists and insurers act on every prediction.
• Models are served on CPU in a serverless setup for cost, so they must deliver quality without GPU inference.
• Off-the-shelf detectors add friction - the YOLO family runs into Ultralytics commercial licensing thresholds.
• In-house baselines had plateaued despite earlier custom training and self-supervised work.

The team had already used Lightly's open-source SSL framework with measurable gains, so LightlyTrain was a natural place to start.

Adravision adopted LightlyTrain to train DINOv3-based object detectors for dental radiographs, then benchmarked them against the approaches the team had used or built: multiple YOLO versions, Faster R-CNN, RF-DETR (Roboflow), and custom models with various backbones. The LightlyTrain models came out ahead.

Choosing the right backbone

The team compared convolutional and Vision Transformer backbones with DINOv3. The ViT backbones delivered substantially better detection quality on Adravision's datasets, so the team selected them and accepted the modest speed tradeoff - kept manageable by their cost-optimized serverless serving. The move also built on the team's earlier LightlySSL work, making it easy to fold into their existing pipeline.

‍

With LightlyTrain, Adravision built dental radiograph object detection models that:

• match or outperform their previous YOLO, Faster R-CNN, RF-DETR, and custom baselines,
• meet the accuracy bar their clinical and insurance users demand,
• run cost-efficiently on CPU in a serverless architecture, with no GPU required,
• fit cleanly into a workflow already built on Lightly's self-supervised tooling.

Object detection is now the strongest part of the team's stack. Next, Adravision plans to use DINOv3 pretraining in LightlyTrain on ~100,000 unlabeled radiographs to build a domain-specific dental foundation model.

‍