Medical Physics Seminar - Cari Whyne & Michael Hardisty

Authors: Michael Hardisty (speaker), Geoff Klein, Arjun Sahgal, Anne Martel, Cari Whyne (speaker)

An assessment of spinal instability is an essential component of decision making in the multidisciplinary treatment of spinal metastases. Current clinical use of the Spinal Instability Neoplastic Score (SINS) requires manual calculation of the SINS elements and it has been reported that experience has a significant impact on the reliability of this score. We have developed an automated pipeline for the prediction of the SINS elements in the metastatic spine using deep learning based on input CT data. Further extension of this work has been applied to understand the progression of osteosarcopenia in patients with advanced cancer.  A novel multitask architecture with a ResNet-50 convolutional backbone is used to generate multiple output feature maps, at the whole spine level and for each vertebra which are combined to yield the elements used in SINS. The pipeline uses this model to label and segment the vertebrae and spine musculature, quantify bone density and muscle quality, identify and calculate the % of vertebral metastatic involvement (osteolytic/osteoblastic), identify involvement of the posterior elements, estimate the % collapse of fractured vertebrae, and calculate spinal malalignment. Instance segmentation of the vertebrae is accomplished using a composite loss to train the model end to end, which yields a useful feature representation that is used for downstream tasks. To quantify bone lesions, osteolytic and osteoblastic disease are each individually quantified using a histogram-based approach. Using an ensemble method, sagittal and coronal spinal alignment are calculated from the multitask ResNet vertebral location predictions, where the centroid of each vertebra and the planes of its endplates are used to make angle calculations based on the local curvature of the spine. To calculate vertebral body collapse, our pipeline predicts what the intact volume of a fractured vertebra should be through interpolation based on the volume of the adjacent vertebrae and then calculates the loss in volume. Finally, the detection method uses vertebral specific feature maps generated from the ResNet50 backbone with additional convolutional layers trained to classify vertebrae as having unilateral or bilateral involvement of the posterior and lateral elements. Quantitative imaging biomarkers calculated from standard of care imaging has the potential to improve fracture risk prediction and better understand long term musculoskeletal health in this patient population.

A video will be available after the presentation.