Uncertainty-guided test-time optimization for personalizing segmentation models in longitudinal medical imaging.

Area of Science:

• Artificial Intelligence in Medical Imaging
• Machine Learning for Healthcare
• Computer-Aided Diagnosis

Background:

• Accurate segmentation of longitudinal medical scans is crucial for patient monitoring and treatment.
• Current personalized models often use fixed training durations and lack optimal stopping criteria, especially without validation labels.

Purpose of the Study:

• To develop an uncertainty-guided test-time optimization (TTO) framework for dynamic, patient-specific personalization of segmentation models.
• To enable validation-free stopping criteria based on predictive uncertainty.

Main Methods:

• Personalized a generalized segmentation model using patient-specific prior scans and test-time optimization.
• Utilized Monte Carlo Dropout (TTO-MCD) and Deep Ensembling (TTO-DE) to estimate voxel-wise predictive uncertainty.
• Evaluated on pancreas, liver, and head-and-neck tumor datasets, comparing against a fixed-epoch baseline (Pre-TTO).

Main Results:

• TTO methods significantly outperformed Pre-TTO and unpersonalized baselines across multiple segmentation metrics (DSC, HD95, MSD, LPS).
• Statistically significant improvements were observed for pancreas and liver datasets (p < 0.05).
• TTO-MCD and TTO-DE achieved high performance without requiring test-time labels.

Conclusions:

• Uncertainty-guided TTO offers a robust, validation-free approach for optimizing patient-specific segmentation models in longitudinal imaging.
• The framework enhances segmentation quality across various modalities and anatomical targets.
• This method supports clinical deployment of personalized AI in medical imaging.