Verified Tracking Paradigm Improves Lesion Segmentation in CT Scans
A new Verified Tracking paradigm for tumor lesion tracking across serial CT scans is proposed, combining registration-based prompt verification with longitudinal context to improve segmentation accuracy. The framework uses early spatial prompt fusion and latent temporal difference weighting, achieving up to 4.5 Dice improvement through large-scale synthetic pretraining.
Key facts
- Tracking tumor lesions across serial CT scans is essential for oncological response assessment.
- End-to-end trackers offer high automation but no opportunity to correct silent tracking failures.
- Decoupled registration-segmentation pipelines permit user verification but discard prior lesion appearance.
- The proposed Verified Tracking paradigm involves a clinician verifying a registration-proposed prompt.
- The model leverages the verified prompt alongside baseline lesion appearance to resolve ambiguities.
- The framework combines early spatial prompt fusion with latent temporal difference weighting.
- Large-scale synthetic pretraining is essential for exploiting longitudinal context.
- The method improves performance by up to 4.5 Dice.
Entities
—