Scaphoid fracture detection

This algorithm detects scaphoid fractures on frontal view x-rays. It applies two convolutional neural networks consecutively: a segmentation network localizes the scaphoid and passes the relevant region to a detection network for fracture detection.

Specifications

Input:

  • Anterior-posterior or posterior-anterior view x-ray of the hand, wrist, or scaphoid. The hand should be pointing upwards and should not be rotated more than 45 degrees. Only DICOM or MHA/MHD files are supported, since the algorithm uses the pixel spacing information to normalize the size of the scaphoid.

Output:

  • Scaphoid fracture probability [0-1].
  • Saliency map that indicates which regions in the x-ray were influential to the prediction of the algorithm. These regions correlate with the location of the fracture. A heat map colour coding is used to indicate the importance of the regions: colder (blue-green) and warmer (yellow-red) colours respectively indicate a low and high importance. The saliency map is projected as an overlay on the x-ray (the opacity can be controlled with a slider). The detected location of the scaphoid is indicated by a yellow bounding box.