An automated fracture detection from pelvic CT images with 3-D convolutional neural networks

Abstract

The demand for an automatic bone fracture detection in the emergency section of the hospitals is high for quick diagnosis while maintaining the quality. Previous studies on fracture detection with computed tomography (CT) images or X-ray images have a performance limitation because those methods are based on 2-D image analysis and cannot consider the 3-D internal structure of pelvic bones. This study proposes an automated bone fracture detection from 3-D CT images. Firstly, it introduces a new 3-D annotation method of fractures (called 3-D surface annotation). By using 3-D shape data of pelvic surfaces, it decreases the annotation load significantly. The proposed method estimates the degree of fracture for each point on the pelvic surface. The degree is estimated by 3-D convolutional neural networks (CNN) using 3-D distribution of CT values inside the pelvic surface. The proposed method was validated by using 103 subjects. The accuracy, precision, recall, and specificity for the test data were 69.5%, 61.1%, 56.4%, and 77.7%, respectively.