UNet based automatic femur segmentation with few annotated data for bone fracture prediction: from pre-processing to segmentation assessments
Résumé
Bone metastases are quite frequently found in cancer patients, as bone is the third most common site of metastasis, behind lung and liver. Prostate and breast cancer are responsible for up to 70% of the skeletal metastases. One of the most common being pathological fractures, which are induced by the unregulated bone remodelling caused by the metastases. They often occur in the long bones (i.e. femurs) or the vertebrae, with catastrophic implications. Clinical score exist, like the Mirels score for the femurs, to assess the risk of fracture, but are insufficient to obtain a solid prediction, due to the lack of specificity of the methods, the limitations linked to quantifying pain, and the absence of precise differentiation of the score between lesions located in a similar area. Those methods can be insufficient, and the need for early diagnosis and treatment is important to improve their efficiency. Pathological fractures in particular are difficult to predict, therefore enhancing the ability for an early diagnosis by a numerical simulation can help to greatly decrease the occurrence of such fractures. To develop a numerical model, the main objective of this work is to implement an accurate and automatic femur segmentation using deep learning which enables fracture simulation and prediction.