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

Spinal cord tumors lead to neurological morbidity and mortality. Being able obtain morphometric quantification (size, location, growth rate) of the tumor, edema, cavity can result in improved monitoring treatment planning. Such requires segmentation these structures into three separate classes. However, manual three-dimensional is time consuming, tedious prone intra- inter-rater variability, motivating development automated methods. Here, we tailor a model adapted spinal tumor task. Data were obtained from 343 patients using gadolinium-enhanced T1-weighted T2-weighted MRI scans with cervical, thoracic, and/or lumbar coverage. The dataset includes most common intramedullary types: astrocytomas, ependymomas, hemangioblastomas. proposed approach cascaded architecture U-Net-based models that segments two-stage process: locate label. first finds generates bounding box coordinates. images are cropped according this output, leading reduced field view, which mitigates class imbalance. then segmented. cavity, edema (as single class) reached 76.7 ± 1.5% Dice score alone 61.8 4.0% score. true positive detection rate was above 87% for cavity. To best our knowledge, fully automatic deep learning segmentation. multiclass pipeline available Cord Toolbox (https://spinalcordtoolbox.com/). It be run custom data on regular computer within seconds.