A Deep Learning Based Framework for 3D Reconstruction of Femur from QCT Images

Abstract

Deep Learning (DL) based techniques have been proven to be very effective in medical image segmentation and reconstruction of 3D anatomies of a human body. The success of DL methods primarily depends on extensive and accurately annotated datasets. In this research, we propose a semi-supervised deep learning method that we call SSDL utilizing a CNN-based 3D U-Net model for femur segmentation from sparsely annotated Quantitative Computed Tomography (QCT) slices. We have focused only on annotating QCT slices at the proximal end of the femur, forming ball and socket joint with acetabulum for precise segmentation. Using our proposed framework, we generate a segmenting binary mask to segment the femur accurately. Using our proposed framework, a modified DICOM file integrated with the original metadata can be generated, preserving all the required information. We have employed polynomial spline interpolation for 3D reconstruction along with an Island Removal algorithm to eliminate noises, if there is any. The performance of segmentation and 3D reconstruction have been evaluated both qualitatively and quantitatively. Our approach can achieve a Dice Similarity Coefficient of 91.8% for unseen patients and 99.2% for validated patients. Additionally, we have obtained an average Relative Error of 6.61% and 12.08% for volume and surface area, respectively. The proposed approach demonstrates its effectiveness in accurately segmenting and reconstructing the 3D femur from QCT slices.