Automatic multiclass intramedullary spinal cord tumor segmentation on MRI with deep learning

Area of Science:

• Neuroimaging
• Medical Artificial Intelligence
• Oncology

Background:

• Spinal cord tumors cause significant neurological disability and mortality.
• Accurate morphometric quantification of tumors, edema, and cavities is crucial for effective monitoring and treatment planning.
• Manual segmentation is time-consuming, subjective, and prone to variability, necessitating automated solutions.

Purpose of the Study:

• To develop and validate a fully automatic deep learning model for segmenting spinal cord tumors, edema, and cavities.
• To improve the efficiency and reduce variability in the analysis of spinal cord tumor imaging data.
• To provide a tool for enhanced clinical decision-making in spinal cord tumor management.

Main Methods:

• A cascaded U-Net-based deep learning architecture was developed for a two-stage segmentation process (locate and label).
• The model was trained and evaluated on 343 patient MRI scans (T1- and T2-weighted) including astrocytomas, ependymomas, and hemangioblastomas.
• A preprocessing step cropped images to a reduced field of view to mitigate class imbalance.

Main Results:

• The model achieved a Dice score of 76.7 ± 1.5% for segmenting tumors, edema, and cavities (as a single class).
• Segmentation of tumors alone reached a Dice score of 61.8 ± 4.0%.
• The true positive detection rate for tumor, edema, and cavity exceeded 87%.

Conclusions:

• This is the first fully automatic deep learning model for multiclass spinal cord tumor segmentation.
• The developed pipeline, available in the Spinal Cord Toolbox, offers rapid and accurate analysis on standard hardware.
• The automated approach has the potential to significantly advance spinal cord tumor monitoring and treatment planning.