Human Bone Assessment: A Deep Convolutional Neural Network Approach

Abstract

The cancerous or fracture bone reduces the capability to perform our daily routine. Therefore, bone diagnosis is required as soon as possible. Bone diagnosis is performed by a doctor using CT scan, X-ray, magnetic resonance imaging (MRI), and digital imaging and communications in medicine (DICOM) image. The diagnosis depends on the expertness of a doctor. The computer-assisted diagnosis (CAD) is a reliable and efficient tool for the diagnosis of a human bone. In the past researches, bone is either classified into a fracture or non-fracture. Some of the researches also classify cancerous vs. healthy bone. The proposed approach classifies the healthy, fracture, and cancerous bone using a deep convolutional neural network (CNN) model. A deep CNN model automatically learns features from the image. Training of a deep CNN model with small datasets may lead to overfitting. Therefore, we have trained our model with 40,800 bone images. The proposed BResNeXt is based on a residual deep CNN topology. The BResNeXt is performing outstanding by producing an overall accuracy of 94.54%. The class-wise precision of the model is 96% (cancerous), 90% (fracture), and 99% (healthy). The F1-score for cancerous, fracture, and healthy bone is 93%, 94%, and 96%, respectively. This novel research will open a path to perform a diagnosis of cancerous, fracture, and healthy bone on a single platform.