Résumé: Chest rib fracture can be regarded as a kind of small objects to be detected with complex shape and large similarity to the surrounding background. Fatigue caused by work intensity is the main reason why radiologists’ detection efficiency and accuracy of rib fracture decrease over time. This work proposes an automatical detection method of rib fractures in chest CT image based on a centernet network with heatmap pyramid structure. Firstly, a hierarchical fusion hourglass network is constructed to perform feature extraction, and the multi-branch residual blocks and fusion strategy in it also play a positive role in improving the feature extraction ability. Secondly, a heatmap pyramid structure which can utilize multi-scale information is adopted to generate more accurate corner information. In addition, a non-local dual spatial attention module is designed to reduce the misclassification of corner and center points. Experimental results show that the proposed network achieves an accurate detection of rib fracture with a mean Average Precision (mAP) of more than 89%.

Langue: Anglais