Wrist Analysis in radiographs
Detection of Wrist Fractures
- We developed the first fully automated system to analyse the wrist in the two standard views (i.e. PA and LAT).
- The system achieved an encouraging fracture detection rate, with an Area under Receiver Operating Characteristic Curve (AUC) of 0.93 from LAT view, of 0.95 from PA view, and of 0.96 from both views combined, in cross-validation experiments on a dataset of 1010 cases, half of which with fractures.
- The system comprises a global searcher locating the distal radius and a local searcher segmenting its contour. The global search was performed by Random Forest Regression-Voting (RFRV) while the local search was performed by Random Forest Regression-Voting Constrained Local Model (RFRV-CLM).
- The resulting automated annotation gives information on the position, orientation and scale of the distal radius accurately so that registered radiographic patches can be cropped.
- A Convolutional Neural Network (CNN) was trained per view from scratch on registered patches.
- The decisions from both views are combined by averaging.
