Neural networks for bone fracture classification

Developing technologies are rapidly emerging every day in different fields, especially in medical environment. In particular, among these technologies, computer vision based products have really helped doctors in the last few years. Suspected fractures are among the most common reasons for patients to visit emergency departments, and X-ray imaging is the primary diagnostic tool used by clinicians to assess patients for fractures. Sometimes the size of fractures is not significant and could not be detected easily and missing a fracture in a radiograph often has severe consequences for patients, resulting in delayed treatment and poor recovery of function. Therefore, effective and intelligent systems should be designed. We wanted to design a tool able to help doctors in diagnosis of bones’ fractures, that, as we said, are often hard to recognise in X-ray to the naked eye. Given the complexity of certain fractures, we thought that using a standard classification approach wouldn’t be optimal for this problem. That’s why we decided to implement a solution based on a neural network. Deep Learning and neural networks are being praised as a major disruptive innovation with the potential to transform most industries and today they achieve outstanding performances on many important problems, in particular in computer vision, speech recognition, and natural language processing. In the conventional approach to programming, we tell the computer what to do, breaking big problems up into many small, precisely defined tasks that the computer can easily perform. By contrast, in a neural network we don’t tell the computer how to solve our problem. Instead, it learns from observational data, figuring out its own solution to the problem at hand. Neural networks may have many implementations depending on the field of application. As we’re going to apply them in computer vision, the most common class used is Convolutional Neural Network (CNN), a technology that lately have reached or even surpassed human precision in a lot of tasks, for example handwritten digits’ recognition. That’s why we decided to use a CNN to classify different type of fractures in proximal femurs, one of the most common fracture nowadays. There are three main groups of fractures in in this bone, and each of this group can be divided in subgroups, depending from the area involved and from the shape of the fracture. The dataset is composed of approximately 2500 images, divided in different classes, that will be increased with data augmentation techniques in order to have a proper number of images. The labelling was done by doctors from the CTO (Traumatology Orthopaedic Centre) of Turin under the supervision of a senior specialised orthopaedic surgeon. The thesis is divided into three main parts: in the first part we discussed briefly all the main aspects of Deep Learning, starting from the basis until the latest introductions in this technology, in order to better understand why we have chosen this approach to solve our classification problem. In the second part we reviewed and commented pros and cons of the existing approaches to this type of problem. The third part is the developed system, that comprises two principal stages. In the first stage, the original images of the fractures are processed using different image processing techniques in order to detect their location and shapes. The second stage is the classification phase using neural networks.