Semantic Segmentation of Femur Bone from MRI Images of Patients with Hematologic Malignancies

Femur bone marrow MRI images may provide more information for better characterizing hematologic malignancies and understanding their prognostics compared with blind bone marrow biopsies and aspirates. However, the interpretation of femur bone marrow MRI images is time-consuming and intervariable among different physicians. To develop a computer-aided diagnosis system for hematologic malignancies using femur bone marrow MRI images, we propose a fully automatic method for femur bone segmentation using deep learning. Five classical pretrained networks such as U-Net and Segnet based on a backbone of Resnet50 network are used for femoral bone segmentation. We conducted an experimental study on 200 cases with hematologic malignancies. Through training using 149 patients' T1-enhanced MRI images, we obtained an average Dice coefficient of 0.907 using Segnet based on a validation using 38 patients' T1-enhanced MRI images.