Thyroid nodule segmentation on dynamic contrast-enhanced magnetic resonance imaging based on Spatial -Temporal information fusion

Motivation: Automatic segmentation of thyroid nodules on DCE-MRI is vital for improved characterization and diagnosis. Traditional segmentation methods often rely on a static phase , sacrificing valuable spatio-temporal information inherent in DCE-MRI. Goal(s): This study aimed to develop a deep-learning model that integrates spatial and temporal pharmacokinetic information in DCE-MRI for improved segmentation. Approach: We designed a novel deep-learning workflow that integrates spatial features and temporal pharmacokinetic information from DCE-MRI, accounting for individual differences in contrast enhancement and morphology, to overcome the limitations of methods focusing solely on spatial information. Results: Our model outperformed eight classic segmentation networks, confirming its effectiveness in thyroid nodule segmentation. Impact: The success of our model may inspire further research into advanced deep-learning architectures that harness individual intensity variations, morphological priors, and temporal pharmacokinetic information. This general approach could extend beyond DCE-MRI to encompass other medical imaging modalities containing temporal information.