TopoPult-SSL: AI segments meibomian glands without dense masks, beats benchmarks
New self-distilled model uses cheap clinical priors to achieve 0.716 Dice across devices.
TopoPult-SSL tackles a key problem in medical AI: domain shifts between imaging devices. Every new device changes image characteristics, but collecting dense gland masks is expensive and slow. The authors observe that cheap clinical signals—eyelid outlines, Pult grades, morphometric ratios—are routinely recorded and can serve as weak priors. Their two-stage framework first adapts a source-trained model to a target device without any target gland masks, using four weak-prior anchors driven by eyelid masks and metadata. This stage alone achieves Precision 0.694 on the CAMG dataset (100 meibography images from a different device), far outperforming generalist segmenters like SAM (0.30–0.34, p<0.001).
In Stage 2, when a small number of target gland masks become available, the model applies supervised self-distillation: multiple Stage-1 teachers (each specialized via a different weak prior) are distilled into one compact student network. This student surpasses both the ensemble of teachers (Dice 0.720) and the state-of-the-art UA-MT (0.710), achieving 0.716±0.006 (best 0.726) in a single forward pass. Code and reproducibility scripts are released. The work demonstrates how to leverage routinely collected clinical data to reduce the annotation burden for cross-device deployment in ophthalmology.
- Stage 1 uses weak clinical priors (eyelid outlines, Pult grades) to achieve Precision 0.694 without gland masks, vs. 0.30–0.34 for SAM/MedSAM.
- Stage 2 self-distills multiple teachers into one student, reaching Dice 0.716±0.006 on the MGD-1k→CAMG cross-device benchmark.
- TopoPult-SSL requires no dense gland contouring on new devices, enabling rapid deployment in clinical settings.
Why It Matters
Eliminates need for expensive gland annotations in new imaging devices, accelerating AI adoption in ophthalmology clinics.