The effect of deep convolutional neural networks on radiologists' performance in the detection of hip fractures on digital pelvic radiographs

• Published in: European journal of radiology 2020-09, Vol.130, p.109188, Article 109188
• Main Authors: Mawatari, Tsubasa, Hayashida, Yoshiko, Katsuragawa, Shigehiko, Yoshimatsu, Yuta, Hamamura, Toshihiko, Anai, Kenta, Ueno, Midori, Yamaga, Satoru, Ueda, Issei, Terasawa, Takashi, Fujisaki, Akitaka, Chihara, Chihiro, Miyagi, Tomoyuki, Aoki, Takatoshi, Korogi, Yukunori
• Format: Article
• Language: English
• Subjects: Aged; Aged, 80 and over; Area Under Curve; Deep Learning; Diagnosis, Computer-Assisted; Female; Femoral Fractures; Hip Fractures - diagnostic imaging; Humans; Image Processing, Computer-Assisted - methods; Magnetic Resonance Imaging - methods; Male; Middle Aged; Neural Networks, Computer; Pelvis - diagnostic imaging; Radiographic Image Enhancement - methods; Radiography; Radiography, Abdominal - methods; Reference Standards; ROC Curve; Tomography, X-Ray Computed - methods
• ISSN: 0720-048X; 1872-7727
• DOI: 10.1016/j.ejrad.2020.109188

•We developed an automated deep learning system for detecting hip fractures on radiographs using CT and MRI as a gold standard annotated by radiologists trying to avoid the omission of false negative cases from the onset of the study.•We also evaluated the diagnostic performance of human readers with and without DCNN output.•DCNN output improved the diagnostic performance including the experienced readers. The purpose of our study is to develop deep convolutional neural network (DCNN) for detecting hip fractures using CT and MRI as a gold standard, and to evaluate the diagnostic performance of 7 readers with and without DCNN. The study population consisted of 327 patients who underwent pelvic CT or MRI and were diagnosed with proximal femoral fractures. All radiographs were manually checked and annotated by radiologists referring to CT and MRI for selecting ROI. At first, a DCNN with the GoogLeNet model was trained by 302 cases. The remaining 25 cases and 25 control subjects were used for the observer performance study and for the testing of DCNN. Seven readers took part in this study. A continuous rating scale was used to record each observer's confidence level. Subsequently, each observer interpreted with the DCNN outputs and rated them again. The area under the curve (AUC) was used to compare the fracture detection. The average AUC of the 7 readers was 0.832. The AUC of DCNN alone was 0.905. The average AUC of the 7 readers with DCNN outputs was 0.876. The AUC of readers with DCNN output were higher than those without(p