LinGuinE: Longitudinal Guidance Estimation for Volumetric Tumour Segmentation
New PyTorch framework achieves state-of-the-art segmentation across 456 longitudinal studies with no longitudinal training.
A research team led by Nadine Garibli has published a new AI framework, LinGuinE (Longitudinal Guidance Estimation), designed to solve a critical gap in medical imaging: tracking tumors over time. Current methods typically produce single-timepoint masks, lack lesion correspondence, and offer limited control for radiologists. LinGuinE addresses this by combining image registration with guided segmentation, allowing a single radiologist prompt to generate volumetric masks and track individual lesions across every scan in a longitudinal study. This is vital for accurate radiotherapy planning and assessing how a tumor responds to treatment.
The framework is built in PyTorch and is uniquely flexible; it is temporally direction-agnostic, requires no training on longitudinal data, and can repurpose any existing registration and semi-automatic segmentation algorithm. In evaluations across four datasets comprising 456 longitudinal studies, LinGuinE achieved state-of-the-art segmentation and tracking performance, with minimal degradation in accuracy as time between scans increased. The team has released their code and a public benchmark, which will significantly accelerate future research in this underexplored but crucial area of computational oncology.
- Framework combines image registration & guided segmentation for lesion-level tracking from a single prompt.
- Achieved state-of-the-art performance on 456 longitudinal studies across four datasets with minimal temporal degradation.
- Requires no training on longitudinal data and is compatible with any existing registration/segmentation algorithm.
Why It Matters
Enables precise, efficient tracking of tumor growth and treatment response, directly improving radiotherapy planning and patient outcomes.