Tailoring Loss to Boost Vertebra Centroid Localization and Classification in Sagittal Spinal Radiographs

Cliodhna Gartland, Fjorda Koromani, John Healy, Gennady Roshchupkin, Fernando Rivadeneira, Stephen J. Redmond

Keywords: Deep Learning, Loss, Spine, Radiographs

TL;DR: We use tailored loss functions to deal with limitations in our dataset and boost performance when localising and classifying the vertebra centroids from spinal radiographs.

Abstract: Osteoporotic vertebral fractures (OVFs) increase the risk of future fractures, morbidity, and mortality. However, manual interpretation of spinal radiographs is time-consuming and challenging. To address this, we propose an automated tool for vertebral centroid localization and classification in thoracolumbar sagittal spinal radiographs with varying fields-of-view, automating the initial input required for currently available semi-automated diagnostic tools and enhancing the efficiency of OVF assessment. To guide our model in its learning, we tested four loss functions to encourage focus on the vertebral centroid locations and to tailor our model to deal with limitations in our dataset, such as unlabeled visible vertebrae, that will aid in its generalizability. Our best performing model achieved a vertebral identification accuracy of 95.9% and a centroid localization RMSE for all correctly classified vertebrae of 17.08 pixels. Future work will leverage the outputs of this model for vertebral fracture detection in a fully-automated pipeline.

Submission Number: 125