Proximal femur fracture detection and classification using deep learning

Proximal femur fractures and dislocation are one of the most serious fractures as they can lead to mortality which are more common in elderly patients, and to limping, inability to work, to practice sport activity in young adults patients if not diagnosed accurately and in real time. Artificial intelligence and its sub-categories, machine learning and deep learning shows a great work when merged in healthcare applications in sustaining doctors to diagnose the diseases accurately and timely. The main objective of this thesis is to detect the various proximal femur fractures types on x-rays using the instance segmentation task with the polygon annotation. For this thesis, a dataset of 756 X-ray images was used. The images were manually annotated using polygon annotation in order to perform the instance segmentation task with the supervison of a consultant and specialist orthopedic doctor. The labels used to annotate the dataset were related to the fracture type (neck, dislocation, grater-trochanter, lesser trochanter, intertrochanteric and subtrochanteric) and normal the ones without any fracture. The dataset was splitted into 70% for training and 30% for validation. In order to test the models performance, the models were given unseen before X-ray images. The models used in order to perform the instance segmentation task were YOLOv5, YOLOv7, YOLOv8, Mask R-CNN (Detectron2) and PointRend. According the results, all five models were able to detect the presence of a proximal femur fracture and classify it. The Mean Average Prediction (mAP) values of Intersection over Union (IoU) threshold of 0.5 (mAP50) for boxes were 79.5%, 79.2%, 81.6%, 73.82% and 73.67% for YOLOv5, YOLOv7, YOLOv8, Mask R-CNN (Model3) and PointRend (Model4) respectively. While, the Mean Average Prediction values of Intersection over Union (IoU) threshold of 0.5 (mAP50) for masks were 80%, 78.8%, 81.7%, 74.05% and 74.85% for YOLOv5, YOLOv7, YOLOv8, Mask R-CNN (Model3) and PointRend (Model4) respectively.