Semi-Supervised Learning Based Femur Segmentation from QCT Images

Jamalia Sultana, Mashiyat Nayeem, Mahmuda Naznin, Tanvir R. Faisal

Abstract: Segmentation of femur in QCT image is always challenging due to its complex ball and socket joint with acetabulum. Recently, deep learning based techniques have been used in image segmentation. However, the success of these methods depends mainly on the accurate annotations which is costly, needs expert opinion, or manual intervention. To overcome these challenges, we propose a semi-supervised learning based approach where we have developed a U-Net based framework for femur segmentation from sparsely annotated Quantitative Computed Tomography (QCT) slices. We annotate only the part of QCT slices at the proximal end of the femur joint. We then integrate the original metadata using instance numbers from QCT files. This semi supervised approach facilitates to work with DICOM medical image data with less annotation. Our framework is cost effective since it saves cost for manual intervention and annotation by medical experts. We have used performance metric Dice Similarity Coefficient (DSC) and found that, we have achieved DSC of 93.7% for unseen patients, and DSC of 99.2% for patients in validation stage which are promising results.