Bone is a common organ for metastasis. Prostate, breast and lung cancers are most probable to spread to the skeleton. Early detection of the pathological abnormalities in bone is critical for precise staging and optimal treatment. As some metastatic lesions come with complicated patterns to detect, specialists encounter difficulties in marking the exact location of them on CT-scan images. Hence, utilizing an automated segmentation mechanism using convolutional neural networks can aid in detecting the metastatic region. Therefore, this project aims to investigate the use of deep learning for recognizing the bone pathological abnormalities due to cancer.

In this Ph.D. project, three different types of bone metastasis (osteolytic, osteoblastic and mixed) in the proximal femur (fig.1) will be taken into account. A set of CT-scan images containing pathological bone will be used to train and test the neural network which will be customized to detect the metastatic region. In combination with a statistical analysis on geometry and bone densities in the proximal femur, different characteristics of detected metastatic lesions such as shape, size and the location could eventually lead to a strength prediction on pathological bone. Based on the strength prediction, specialists could opt for treatment strategies including radiotherapy, surgery, etc. in order to decrease the fracture risk.