TimeLesSeg: Unified Contrast-Agnostic Cross-Sectional and Longitudinal MS Lesion Segmentation via a Stochastic Generative Model

TimeLesSeg introduces a unified deep learning framework for segmenting Multiple Sclerosis lesions that works across different imaging contrasts and with or without temporal data. The model uses stochastic generative techniques and domain randomization to address the fragmentation between cross-sectional and longitudinal segmentation approaches, demonstrating superior performance on multiple datasets.

TimeLesSeg addresses a critical bottleneck in medical imaging AI: the brittleness of current deep learning models when confronted with real-world variability. MS lesion segmentation demands high accuracy because treatment decisions depend on precise lesion quantification, yet existing approaches fracture into separate pipelines for cross-sectional versus longitudinal analysis and struggle when imaging protocols change. This new framework consolidates both pathways into a single neural network, substantially reducing implementation complexity for clinical deployment.

The technical innovation centers on two key mechanisms. First, the model leverages empty mask inputs during training to enable cross-sectional analysis without prior timepoints—a simple but effective solution to a structural mismatch. Second, stochastic morphological deformation of lesion masks synthetically generates plausible lesion evolution patterns, circumventing the chronic scarcity of longitudinal training data in medical imaging. The Gaussian mixture model-based domain randomization exposes the network to intensity variations across different MRI scanners and protocols, a persistent headache for radiologists and AI engineers alike.

For the medical AI sector, this represents tangible progress toward production-ready systems. Clinical institutions increasingly demand models that perform reliably across heterogeneous equipment without retraining. The open-source release amplifies impact, enabling rapid adoption across research centers and hospitals. The superior performance on standardized benchmarks and in-house datasets suggests genuine generalization rather than overfitting to specific domains.

The framework's success with single-modality inputs is particularly noteworthy, as it reduces data acquisition requirements—a significant cost factor in clinical practice. Ongoing validation on larger prospective datasets and integration with existing clinical workflows will determine whether this advances to widespread clinical implementation.

• →TimeLesSeg unifies cross-sectional and longitudinal MS lesion segmentation in a single model, eliminating the need for separate pipelines.
• →Stochastic lesion deformation synthetically generates longitudinal training data, addressing dataset scarcity in medical imaging.
• →Gaussian mixture model domain randomization enables contrast-agnosticism across different MRI scanners and acquisition protocols.
• →Outperforms current state-of-the-art on multiple public and proprietary datasets using single-modality inputs.
• →Open-source release accelerates adoption potential in clinical and research settings.