Automated system for the detection of thoracolumbar fractures using a CNN architecture.

A damage to the spinal cord can affect its natural functioning both temporarily or permanently resulting in loss of sensation and functioning of muscles. Spinal cord injury occurs due to vertebral fractures, especially in thoracolumbar junction. Early detection and treatment of spinal cord injury can help to improve the neurological status. Computed tomography (CT) is one of the effective imaging modalities to visualize the status of vertebrae and make necessary diagnosis. But, these imaging modalities involve both inter and intra observer variabilities which may result in identification of wrong level. Thus, an automated computer aided diagnosis (CAD) system is required to accurately identify the fracture and prescribe correct treatment at the earliest. It is very difficult to precisely segment the vertebra to detect spinal injuries using image processing techniques. Hence, in this paper, we are proposing an automated thoracolumbar fracture detection technique using convolutional neural networks (CNNs) without segmenting the vertebra. The proposed method can efficiently classify the normal and fractured subjects with an accuracy of 99.10%, sensitivity of 100% and specificity of 97.61% using our private dataset (Total image 1120). This novel CAD system can assist the orthopedists in their routine screening.