[Paper Review] Multiclass Spinal Cord Tumor Segmentation on MRI with Deep Learning

Andréanne Lemay, Charley Gros+5 more|arXiv (Cornell University)|Dec 23, 2020

Medical Imaging and Analysis26 references4 citations

TL;DR

This paper presents the first fully automatic deep learning framework for multiclass segmentation of spinal cord tumors, edema, and cavities in MRI using a cascaded U-Net architecture. The method first localizes the spinal cord to crop the region of interest, reducing class imbalance and improving segmentation accuracy, achieving a Dice score of 76.7 ± 1.5% for all three classes and 61.8 ± 4.0% for tumor alone.

ABSTRACT

Spinal cord tumors lead to neurological morbidity and mortality. Being able to obtain morphometric quantification (size, location, growth rate) of the tumor, edema, and cavity can result in improved monitoring and treatment planning. Such quantification requires the segmentation of these structures into three separate classes. However, manual segmentation of 3-dimensional structures is time-consuming and tedious, motivating the development of automated methods. Here, we tailor a model adapted to the spinal cord tumor segmentation task. Data were obtained from 343 patients using gadolinium-enhanced T1-weighted and T2-weighted MRI scans with cervical, thoracic, and/or lumbar coverage. The dataset includes the three most common intramedullary spinal cord tumor types: astrocytomas, ependymomas, and hemangioblastomas. The proposed approach is a cascaded architecture with U-Net-based models that segments tumors in a two-stage process: locate and label. The model first finds the spinal cord and generates bounding box coordinates. The images are cropped according to this output, leading to a reduced field of view, which mitigates class imbalance. The tumor is then segmented. The segmentation of the tumor, cavity, and edema (as a single class) reached 76.7 $\pm$ 1.5% of Dice score and the segmentation of tumors alone reached 61.8 $\pm$ 4.0% Dice score. The true positive detection rate was above 87% for tumor, edema, and cavity. To the best of our knowledge, this is the first fully automatic deep learning model for spinal cord tumor segmentation. The multiclass segmentation pipeline is available in the Spinal Cord Toolbox (https://spinalcordtoolbox.com/). It can be run with custom data on a regular computer within seconds.

Motivation & Objective

• To address the challenge of time-consuming and subjective manual segmentation of spinal cord tumors, edema, and cavities in 3D MRI scans.
• To develop an automated, robust, and efficient deep learning solution for multiclass segmentation of the three most common intramedullary spinal cord tumor types: astrocytomas, ependymomas, and hemangioblastomas.
• To mitigate class imbalance in spinal cord tumor segmentation by focusing on a reduced field of view through initial spinal cord localization.
• To enable accurate morphometric quantification of tumor size, location, and growth rate for improved treatment planning and monitoring.
• To release a user-friendly, computationally lightweight pipeline compatible with standard hardware for clinical and research use.

Proposed method

• A cascaded deep learning architecture is employed, consisting of two stages: spinal cord localization followed by multiclass tumor segmentation.
• The first stage uses a U-Net-based model to detect the spinal cord and generate bounding box coordinates for cropping the region of interest.
• Cropping the MRI volume based on the detected spinal cord reduces the field of view, minimizing computational load and mitigating class imbalance.
• The second stage applies a U-Net-based model to segment three classes: tumor, edema (combined as one class), and cavity, using T1- and T2-weighted gadolinium-enhanced MRI scans.
• The model is trained end-to-end on a dataset of 343 patients with cervical, thoracic, and/or lumbar spinal cord coverage.
• The pipeline is implemented in the Spinal Cord Toolbox and can process new data on standard computers in seconds.

Experimental results

Research questions

• RQ1 Can a cascaded deep learning approach improve multiclass segmentation accuracy of spinal cord tumors, edema, and cavities in MRI compared to existing methods?
• RQ2 To what extent does spinal cord localization and region-of-interest cropping reduce class imbalance and enhance segmentation performance?
• RQ3 Can a fully automatic model achieve high sensitivity and Dice scores for all three target structures (tumor, edema, cavity) across diverse patient scans?
• RQ4 Is the proposed method computationally efficient and deployable on standard hardware without requiring high-end GPUs?
• RQ5 How does the model perform in detecting tumors, edema, and cavities across the three most common intramedullary spinal cord tumor types?

Key findings

• The multiclass segmentation of tumor, edema, and cavity achieved a Dice score of 76.7 ± 1.5%, indicating strong overlap with manual segmentation.
• Tumor-only segmentation reached a Dice score of 61.8 ± 4.0%, demonstrating robust performance despite class imbalance.
• The true positive detection rate exceeded 87% for all three structures—tumor, edema, and cavity—confirming high sensitivity.
• The model’s cascaded design significantly improved segmentation performance by reducing the field of view and minimizing class imbalance.
• The entire pipeline is computationally efficient and can process new MRI data on a standard computer within seconds.
• To the best of the authors’ knowledge, this is the first fully automatic deep learning model for multiclass spinal cord tumor segmentation in MRI.