Convolutional Neural Network Based Femur Stabilization for X-Ray Image Sequences

Abstract

Sequence stabilization of medical images is an important aspect of diagnosis, therapy, joint movement kinematic analysis, and cancer detection. Typically, when image frames are recorded, the body is not rigidly fixed as a result of e.g. respiration, thus the position of its segments may vary. Simple image analysis methods (e.g. gradient based, scale-space based) tend to have problems with discerning the key-points in this specific task, due to large diversity of bone structure and highly visible soft tissue. In this paper, we propose a specialized algorithm for stabilization of femur in a sequence of single plane fluoroscopic images. The method estimates the positions of several easily-detectable femur key-points using gradient-based image analysis methods. For other key-points, which are located in the regions of bone with saliency prohibiting effective detection, we use feedforward Convolutional Neural Network as a position estimator. All the key-point positions are used in a stabilization process performed with the ICP (Iterative Closest Point) algorithm. The overall stabilization accuracy is evaluated for two uncorrelated X-ray image sequences, where manual stabilization (i.e., the results for image alignment performed by a human operator without access to key-points) constitutes the ground truth.