[1M4-CC] Late Breaking Research Session - A

[1M4-CC-03] Estimation of the pelvic fracture degree using 3-D CT images and a 3-D Convolutional Neural Network

Naoto Yamamoto¹, Daisuke Fujita¹, Rashedur Rahman¹, Naomi Yagi², Keigo Hayashi³, Akihiro Maruo³, Hirotsugu Muratsu³, Shoji Kobashi¹ (1. University of Hyogo, 2. Himeji Dokkyo University, 3. Steel Memorial Hirohata Hospital)

In the emergency hospital, physicians and radiologists currently diagnose bone fractures by gazing numerous images. Then, the automatic fracture detection needs to be developed to reduce the burden on physicians and the oversight. This study focuses on the pelvic fracture, which is difficult to identify and tends to cause patients to lie in bed. In conventional methods, only two-dimensional image analysis using simple X-rays or CT images is often utilized as the solution. However, the analysis range is very small, which limits the performance of the method. This study proposes an automatic pelvic fracture detection method from 3-D CT images. Firstly, 3-D shape data representing the bone surface is created from 3-D CT images. Then, the coordinates around the bone surface and CT values are calculated as the 3-D feature vector. The distance from the fracture is also calculated for each voxel, and the fracture probability is assigned to each bone surface simultaneously. The fracture probability is estimated from the features at each bone surface by using a 3-D convolutional neural network (CNN). The proposed method was evaluated on 110 subjects which is divided to 70 training data and 40 validation data, and the results showed that the AUC of the training data was 0.90, and the AUC of the validation data was 0.84. Integrating these results into the 3-D data will bring physicians and radiologists the great reference of the fracture detection.