gms | German Medical Science

Objectives/Interrogation: Distal radius fractures (DRF) are the most common fractures of the human body. The diagnose is traditionally based on clinical examination and x-ray pictures. The primary diagnosis is often carried out by a general practitioner with varying degrees of experience. Sometimes fracture or fracture displacement may be missed. Deciding the appropriate treatment can also cause problems.

The past decades advances in computer sciences have made possible to develop artificial intelligence (AI) algorithms to analyse medical images. This is shown in other studies made with hip, shoulder and chest x-rays.

In this study, we present the results of a convolutional neural network (CNN) for DRF detection.

Methods: Nearly 10 000 posteroanterior and lateral wrist x-ray views were extracted from radiological archive. The x-rays were identified by ICD-10 codes correlating with DRF and wrist bruises (as a control group). The anonymized dataset was divided into a training set, validation set and a held-out test set. The x-rays in training and validation sets were labelled by one hand surgery resident. The held-out test set, that was not used in neural network training, was labelled by three experienced consultant hand surgeons. After all the x-rays were labelled a CNN was trained using the training and validation sets. The results of the CNN were compared to the test set's ground truth set by the three hand surgeons. The development of the CNN was done in co-operation with third parties and hospital's AI experts.

Results and Conclusions: The preliminary results showed that CNN has great potential in DRF detection. The details are opened and discussed in the presentation. The sensitivity and specificity of CNN for DRF detection are promising and further development into a highly efficient tool is possible.

Our preliminary results indicate that with added amount of wrist x-rays and further training a CNN will be a reliable aid in DRF detection. Continuous development, tests and validation in clinical settings are needed before a CNN can be used as a medical aid. The future challenges lie in assessing fracture dislocation and the anatomical changes caused by the fracture. AI algorithms will revolutionise medical image interpretation, from x-ray to 3D imaging, in the near future.