Convolutional neural network for detecting rib fractures on chest radiographs: a feasibility study

Area of Science:

• Radiology
• Medical Imaging
• Artificial Intelligence

Background:

• Chest radiography is standard for rib fracture detection.
• AI application is limited by image quality and multilesion screening challenges.
• Few studies have verified AI performance on multi-center radiographs for rib fractures.

Purpose of the Study:

• Develop a multiple rib fracture detection model using a convolutional neural network (CNN).
• Utilize multi-center and quality-normalized chest radiographs for model training and verification.
• Create a faster and more efficient algorithm compared to existing methods.

Main Methods:

• Adopted an object detection CNN, You Only Look Once v3 (YOLOv3), for model development.
• Utilized 1080 radiographs with rib fractures for training and testing sets.
• Evaluated model performance using Receiver Operating Characteristic (ROC) and Free-Response ROC (FROC) curves.

Main Results:

• Achieved high sensitivity (91.1-92.0%) and precision (81.6%) in testing sets.
• Demonstrated 91.3% sensitivity for whole-lesion detection with a low false-positive rate.
• Showcased case-level accuracy (85.1%), sensitivity (93.2%), and specificity (79.4%) in a joint testing group.

Conclusions:

• The CNN model based on YOLOv3 is sensitive for detecting rib fractures on chest radiographs.
• The model shows significant potential for preliminary screening, reducing missed diagnoses.
• This AI approach can assist radiologists and alleviate workload in fracture detection.