Automatic lymphedema segmentation in T2-STIR MRI using an unsupervised clustering method.
Authors
Affiliations (5)
Affiliations (5)
- Radiology Department, Fondazione IRCCS Ca' Granda Ospedale Maggiore Policlinico di Milano, Via Francesco Sforza 35, 20122, Milano, Italy.
- Politecnico di Milano, Piazza Leonardo da Vinci 32, 20133, Milano, Italy.
- Department of Computer Science, Università degli Studi di Milano, Via Festa del Perdono 7, 20122, Milano, Italy.
- Postgraduate School in Radiodiagnostic, Università degli Studi di Milano, Via Festa del Perdono 7, 20122, Milano, Italy.
- Radiology Department, Aziende Socio Sanitarie Territoriale Fatebenefratelli Sacco, Ospedale Fatebenefratelli e Oftalmico, Piazzale Principessa Clotilde 3, 20121, Milano, Italy. [email protected].
Abstract
To develop and evaluate an unsupervised artificial intelligence (AI)-based method for the automated segmentation and quantitative assessment of fluid content on T2 short tau inversion recovery (STIR) magnetic resonance lymphography (MRL) in patients with lymphedema and lipolymphedema. A heterogeneous cohort of 20 patients with lymphedema or lipolymphedema was retrospectively selected from our clinical database. Image segmentation was performed using a K-means clustering algorithm based on pixel intensity, with single-slice manual segmentation serving as the reference standard. The clustering algorithm was optimized by maximizing the Dice similarity coefficient up to a fixed threshold. The selected clusters were then saved for transfer learning and applied to the test set, where segmentation performance was evaluated. The procedure was then applied iteratively in a multi-slice setting, enabling three-dimensional analysis. The distribution of edema was visualized using color maps and plots depicting its spatial pattern along the limb. Additionally, an interactive interface was developed to overlay graphs from multiple examinations, facilitating longitudinal comparisons. The model effectively segmented lymphedema regions, achieving a satisfactory Dice similarity coefficient of at least 0.8 on the training set compared with the manual reference standard. In the test set, the model achieved a Dice score of 0.74 ± 0.05, demonstrating good agreement with manual annotations. The stacked line plots allowed clear visualization of edema distribution along the limb, providing a volumetric representation of edema and enabling the tracking of longitudinal changes in edema patterns over time. This unsupervised AI-based method shows promise for automated segmentation and quantification of edema in T2-STIR MRL. It offers potential for objective assessment of lymphedema and lipolymphedema, aiding in diagnosis, staging, and treatment monitoring.