Skychain: Neural Network for Bone Abnormalities Recognition

Hello, friends! We have recently made a video about our developers and the neural networks they work on. One of them — Alexey Nazarov — is now ready to share with you more info about the development of his ANN (Artificial Neural Network). He calls it MURA (from words “musculoskeletal radiographs”).

The ANN contains a set of musculoskeletal X-ray images, which were manually marked as “normal” and “abnormal.” Each of these refers to one of the standard seven types of radiologic investigation of the upper limbs: fingers, hand, wrist, elbow, forearm, upper arm and shoulder.

The ANN for Bone Abnormalities Recognition detects fractures, hardware, degenerative joint diseases, and other miscellaneous abnormalities, including lesions and subluxations.

To predict the abnormality, the DenseNet169 model with the ImageNet weights is used. The last fully-connected layer was replaced in order the model predicts only two classes at the output. There is a sigmoid activation function used.

As a loss function, binary cross entropy is taken, Adam with default parameters is applied as an optimization algorithm. The training speed is 0.0001.

Examples of Radiograms on Which the Neural Network is Trained

Alexey gave us some examples of X-ray images he uses for training the ANN.

These examples show normal elbow investigation, investigation of abnormal finger with degenerative changes, abnormal forearm investigation showing operative plate and fixation of radial and ulnar fractures and an abnormal investigation of the shoulder joint with a fracture.

The ANN for Bone Abnormalities Recognition Predicts

Here is the Binary Accuracy Graph of the Skychain Neural Network

For now, the accuracy of the ANN is about 82%.

Alexey says he is going to do fine-tuning of model parameters and to remove noise from images. This will improve the accuracy of the ANN.

Join Skychain on social media: Twitter, Facebook, Telegram

Iva Chernysheva, Marketing Manager