Fracture detection from X-Rays with deep learning
Tutor / directorGarcía Gasulla, Dario
Document typeMaster thesis
Rights accessOpen Access
Bone fractures are a widespread medical condition that involves many people and money in the treatment and tracking of this disease. Spain invests lots of money every year with X-Ray scanners, radiologists, and medical staff to improve patients' treatment. However, the detection of the fracture after taking the X-Ray scan is performed manually by the doctor or radiologist. Nowadays, modern scanners produce digital radiography images with high-quality resolution and detail. Still, we are not taking advantage of these massive amounts of data to improve the current way of diagnosing, when we could apply some Computer Vision techniques to automatize or speed up the process. The advent and constant improvement process of Convolutional Neural Networks and Deep Learning in general during the last years has totally transformed the current state of the art in Computer Vision tasks, for a great variety of problems, presenting new approaches in terms of model architectures, training techniques and also with the development and research frameworks. However, even if the models and the overall process is excellent, Deep Learning techniques rely on the quality and quantity of the data. There are not many public datasets out there for radiological imaging that are large enough to obtain satisfactory results for real-world data, always considering the prediction error risk when we are dealing with a medical domain. As a member of the High-Performance Artificial Intelligence research group from Barcelona Supercomputing Center, we present this project in cooperation with the mutual insurance company Asepeyo, which shared a data set of wrist radiography with us, to let us do some research and analysis about the viability of implementing a Decision Support System to help doctors to detect bone fractures from X-Ray images. We present a pre-processing pipeline system that allows us to create a usable data set from large raw radiography data sets efficiently. Finally, we discuss the performance of several Deep Learning models on the task of analyzing the presence of fractures in an X-Ray image.
SubjectsNeural networks (Computer science), Artificial intelligence, Imaging systems in medicine, Xarxes neuronals (Informàtica), Intel·ligència artificial, Imatgeria mèdica
DegreeMÀSTER UNIVERSITARI EN INTEL·LIGÈNCIA ARTIFICIAL (Pla 2017)