Detecting Bone Fracture Using Transfer Learning

Abstract

Deep learning is a key technology which is used in medical image analysis to enhance informed decision-making. Decision-making models need accuracy for better outcomes. In this chapter, we have presented an application of Transfer Learning (TL) to detect open bone fracture using limited set of images. TL is a kind of self-evolving Deep Learning (DL) technique. In each iteration of TL, the model is improvised based on the knowledge transferred from the previously learned task. One of the research challenges in medical image analysis is the unavailability of large data set publicly. This data set is necessary for deep learning methods. There are many sources of image generation such as X-rays, CT scan, ECG, and Ultrasound. But, these images are not available to researchers due to privacy issue of the patients. In order to overcome this limitation of availability of large data set, we have used augmented data set. A small data set has been augmented to increase the image orientation and count. We have applied DL based Convolution Neural Networks (CNN) on medical images for overcoming the limitation of availability of large data set for training. We have attempted to solve the problem of open bone fracture detection using limited number of images. We have achieved this by applying Speeded Up Robust Features (SURF) extractor on preprocessed radiograph images. SURF is a local feature descriptor and detector. The outcome of SURF extractor is then fed to per-trained models using transfer learning technique. Our proposed system would ease the manual process of identifying the presence and extremity of bone fracture. The proposed approach detects fracture from the given X-ray images with an accuracy of 98.8%. Comparative analysis shows that transfer learning gives significantly comparable results as compared to training model from scratch or even better in some cases. There is a downside of the proposed approach as in some cases, TL is vulnerable to overfitting due to less training data, and in some cases, poor preprocessing might lead to the poor classification of data.