Automated Segmentation, Volume Quantification, and Localization of Intravertebral Spinal Tumors Using Mask-rcnn Networks

Introduction: Intravertebral spinal tumors are a common location for metastatic disease, resulting in major disability and reduction in quality of life for thousands of cancer patients. Additionally, quantification of the volume of these tumors (over time) and localization of tumors can aid in monitoring and surgical planning. However, current approaches towards segmentation mainly focus on spinal cord tumor segmentation. We introduce a machine learning model that utilizes a Mask-RCNN architecture to isolate vertebral bodies and subsequently segment tumors within these 3D volumes on T2-weighted MRI.

Methods: MR images of patients diagnosed with vertebral body tumors (both primary and metastatic) were downloaded and anonymized from our institutions picture archive system. Intravertebral tumors were segmented on 50 MRI series. These segmented series were used to train the Mask-RCNN network which had been pre-trained to isolate vertebral bodies. Training data was augmented with random rotations, contrast changes, and brightness additions. Additionally, 20 MRI-series were used as a hold-out test set to evaluate the network (against ground-truth segmentations) after training. Dice scores, Hausdorff coefficients, average surface distance (ASD), and percentage error were calculated in the segmented tumor volume. All segmentations were cross checked and edited by another clinician to verify accuracy.

Results: The average dice score was 0.71±0.05 (±1 SD). The average Hausdorff coefficient was 8.13±8.27. ASD was 1.93±0.39mm. Average percent volume errors was 8.60±4.31%. No tumors were missed in our test set and all were localized within the correct vertebral bodies.

Conclusion : We present the first deep learning model that automatically segments intravertebral spinal tumors on MRI. We found that the model segmented volumes that were highly accurate ( < 10% error) and did not miss tumors in the test set. This work serves as the basis for objective assessment and monitoring of intravertebral tumors and future research will validate this model on a larger, clinically diverse, dataset.