Deep learning application of vertebral compression fracture detection using mask R-CNN

Abstract

Vertebral compression fractures (VCFs) of the thoracolumbar spine are commonly caused by osteoporosis or result from traumatic events. Early diagnosis of vertebral compression fractures can prevent further damage to patients. When assessing these fractures, plain radiographs are used as the primary diagnostic modality. In this study, we developed a deep learning based fracture detection model that could be used as a tool for primary care in the orthopedic department. We constructed a VCF dataset using 487 lateral radiographs, which included 598 fractures in the L1-T11 vertebra. For detecting VCFs, Mask R-CNN model was trained and optimized, and was compared to three other popular models on instance segmentation, Cascade Mask R-CNN, YOLOACT, and YOLOv5. With Mask R-CNN we achieved highest mean average precision score of 0.58, and were able to locate each fracture pixel-wise. In addition, the model showed high overall sensitivity, specificity, and accuracy, indicating that it detected fractures accurately and without misdiagnosis. Our model can be a potential tool for detecting VCFs from a simple radiograph and assisting doctors in making appropriate decisions in initial diagnosis.

Figure 1

Example of labeled data. Each fracture is labeled with a polygon on the thoracolumbar radiograph based on the MRI results. Each polygon mask contains the x and y coordinates of the polygon mask surrounding it. Each bounding box consists of upper left x, y coordinates, width, and height. The entire labeling process was conducted by two trained orthopedic experts.

Figure 2

Mask R-CNN model architecture.

Figure 3

Backbone network and region proposal network. (a) Backbone network is shown. Feature maps from ResNet are upsampled and resized with 1 x 1 convolution to be concatenated with different scaled feature maps. (b) The region proposal network generates candidate regions for objects by sliding-window, referred to as anchor box on feature maps. Each anchor box performs both classification and bounding box adjustments.

Figure 4

Examples of prediction results of each model. From left to right, actual spine lateral radiograph, ground-truth (expert-labeled fracture masks), and prediction from each model are shown.