Whole anterior visual pathway segmentation from high-resolution MRI using artificial intelligence.
Authors
Affiliations (13)
Affiliations (13)
- Department of Neuroscience, Rehabilitation, Ophthalmology, Genetics, Maternal and Child Health (DINOGMI), University of Genova, Genova, Italy.
- Department of Internal Medicine (DIMI), University of Genova, Genova, Italy.
- Department of Neuroscience, Imaging and Clinical Sciences, Centro I.T.A.B., Università degli Studi G. D'Annunzio di Chieti-Pescara, Chieti, Italy. [email protected].
- Neurocenter of Southern Switzerland, EOC, Lugano, Switzerland. [email protected].
- Department of Health Sciences, Section of Biostatistics, University of Genova, Genova, Italy.
- IRCCS Ospedale Policlinico San Martino, Genova, Italy.
- Princess Máxima Center for Pediatric Oncology, Utrecht, The Netherlands.
- Amsterdam University Medical Centre, Vrije Universiteit, Amsterdam, The Netherlands.
- Radiology Unit, Santi Antonio e Biagio e Cesare Arrigo University Hospital, Alessandria, Italy.
- Neurocenter of Southern Switzerland, EOC, Lugano, Switzerland.
- Faculty of Biomedical Sciences, Università della Svizzera Italiana, Lugano, Switzerland.
- Department of Neuroscience, Imaging and Clinical Sciences, Centro I.T.A.B., Università degli Studi G. D'Annunzio di Chieti-Pescara, Chieti, Italy.
- Istituto Nazionale di Fisica Nucleare, Genova, Italy.
Abstract
Manual segmentation of the whole anterior visual pathway (aVP) from high-resolution magnetic resonance imaging (MRI) is time-consuming and prone to inter-rater variability. We developed and validated a fully automated deep learning framework, "aVP-seg," to perform rapid, multiclass segmentation of the optic nerves, chiasm, and optic tracts in healthy volunteers and multiple sclerosis (MS) patients. We developed and validated a cascaded two-stage three-dimensional convolutional neural network (principal segmentation + refinement) for automated multiclass segmentation of the aVP from 0.6-mm isotropic three-dimensional constructive interference in steady state (CISS) MRI. The model was trained and evaluated in 34 healthy controls and 46 MS patients. Ground truth was derived from manual segmentations by two expert radiologists. Spatial agreement metrics included Dice similarity coefficient (DSC), 95th percentile Hausdorff distance (HD95), and volumetric similarity. Agreement with the ground truth for the whole aVP was high (DSC 0.86 ± 0.03, mean ± standard deviation; 95% confidence interval (CI) 0.85-0.86). Boundary alignment was strong (HD95 1.18 mm ± 0.54; 95% CI 1.06-1.30) and volumetric similarity was high (0.96 ± 0.04; 95% CI 0.95-0.97). Accuracy was consistent for the left and right optic nerves (DSC 0.85-0.86 ± 0.05-0.04) and chiasm (DSC 0.83 ± 0.09), but lower for the left and right optic tracts (DSC 0.74-0.75 ± 0.07-0.07). The aVP-seg provided accurate, automated multiclass segmentation of the whole aVP from high-resolution CISS MRI. This tool may standardize and accelerate the extraction of quantitative biomarkers of aVP integrity in neuro-ophthalmic conditions. Automated multiclass segmentation of the entire anterior visual pathway enables standardized and reproducible preparation of MRI data for quantitative analysis. This approach facilitates future assessment of optic pathway involvement in MS and other neuro-ophthalmic disorders. aVP-seg enabled fully automated segmentation of the entire anterior visual pathway from high-resolution CISS MRI data. Automated segmentation reduces processing time and operator-dependent variability. aVP-seg shows robust performance across both healthy subjects and MS patients.