Vertebral compression fracture detection using imitation learning, patch based convolutional neural networks and majority voting

Abstract

Vertebral compression fractures often go clinically undetected and consequently untreated, resulting in severe secondary fractures due to osteoporosis, and potentially leading to permanent disability or even death. Automated detection of vertebral compression fractures (VCF) could assist in routine screening and followup of incidentally scanned patients, thereby mitigating secondary fractures later. A novel fully automated method for the detection of VCF in 3D computed tomography (CT) of the chest or abdomen is presented in this work. It starts with 3D localisation of thoracic and lumbar spine regions using deep reinforcement learning (DRL) and imitation learning (IL). Six different 3D bounding boxes are generated by the localisation step, achieving an average Jaccard Index (JI)/ Dice Coefficient (DC) of 74.21%/84.71%, and detection accuracy of 97.16 % using 3 different CNN architectures. The localised region is then split into 2D sagittal slices around the coronal centre. Each slice is further divided into patches, on which convolutional neural networks (CNNs) are trained to detect VCF. Four different CNN architectures, namely 3 layered, 6 layered and transfer learning (TL) using VGG16 and ResNet50, were experimented with. The best performing architecture turned out to be the 6 layered CNN. Aggregation is performed on the VCF detection in the 2D Patches extracted from individual bounding boxes, followed by majority voting to arrive at the final decision on the status of VCF for a given patient. An average three-fold cross validation accuracy of 85.95%, sensitivity of 88.10%, specificity of 84.20% and F1 score of 85.94% were achieved on chest images using 6 layered CNN on chest images from 308 patients. An average five-fold cross validation accuracy of 86.67%, sensitivity of 88.13%, specificity of 85.02% and F1 Score of 87.04% were achieved on abdomen images from 168 patients with the 6 layered CNN.

Keywords

• 3D localisation
• VCF detection
• Deep reinforcement learning
• Imitation learning
• Convolutional neural networks
• Majority voting