Abnormality Detection in Humerus Bone Radiographs Using DenseNet

Abstract

Treating of injuries and broken bones through reading musculoskeletal radiographs requires a great deal of expertise. It is common that less experienced doctors initially check the radiographs and have a high chance of getting it misdiagnosed. To avoid such misdiagnosis of abnormalities or injury in humerus bone, Deep Learning and Machine Learning algorithms can be applied. Although sophisticated deep learning models have surpassed human capacity under certain computer vision applications, rapid development in the field of medicine has been hampered by a lack of good model applicability and decent marked data, along with other things. This paper seeks to use the model comprehension and visualization methodology to analyze the deep convolution neural network feature removal procedure on the MURA dataset for the identification of anomalies. First, on the selected dataset of humerus radiographs, certain image pre-processing techniques are used to remove variations in size of the image from the radiographs. The following step was to identify the large data as abnormal or normal using the DenseNet-169 architecture. The suggested approach is a reliable technique for classifying bone disorders, according to the findings of the implementation.