Knowledge-based femur detection in conventional radiographs of the pelvis

Abstract

In this paper we present a knowledge-based femur detection algorithm. The algorithm uses femur corpus constraints, Canny edge detection and Hough lines. For optimal femur template placement in the local area we use cross-correlation. The segmentation itself is done with an optimized active shape modeling technique.

Using the knowledge-based technique we have located 95% of the femur shapes of $N=117$ X-rays. From those 83% of the target femur shapes have been segmented successfully (point-to-point error: $\sim 14$ pixels, point-to-boundary error = $\sim 9$ pixels).

Introduction

Osteoporosis is a severe metabolic disease of skeletal system, which leads to a substantial restriction in mobility of patients as well as to skeletal fractures. Beside the vertebral spine and the distal radius, femoral neck is particularly affected by osteoporosis.

The treatment of neck fractures depends on the type of fracture as well as on bone quality. Commonly, the estimation of bone structures of the injured patients is routinely performed by X-rays of the pelvis, if a traumatic injury of the pelvis or the femoral neck is suspected. The diagnosis is done by a radiologist or a specialist in traumatology. The correct diagnosis depends on the quality of the X-ray and on the experience of the physician. A technical support for the requirement of osteoporosis analysis is the automatic localization of femur in order to place the regions of interest on plane radiography.

The major problem is the exact delineation of the femur shape, especially considering the contrast, the gray scaling and the position. Moreover, due to the superposition of overlapping abdominal structures in plane radiography and variations in shape and position of the patient, locating and exact segmentation of the femur is a challenging task. In case of this, experience in interpretation of X-rays is necessary. Manual shape extraction is time consuming, subjective and error prone, so it is not practicable in clinical routine.

The first main challenge is to find automatically the femur structure in a hip overview X-ray. The easiest way is to pose manually a template close enough to the target shape, as mentioned by Behiels et al. [1]. We, however, propose an automatic knowledge-based location technique using an edge-based technique and Hough lines [2], [3]. The optimal location for initializing the template is found by cross-correlation.

After successful template initialization an optimal segmentation technique should be applied. Standard low level segmentation techniques, relying especially only on pixel intensities, often generate incorrect object delineations. One idea to handle this problem is to incorporate a priori information of shape in the segmentation process. Based on that, the most promising method described so far, has been published by Behiels et al. [1]. This method is based on active shape models (ASM) [4], [5] and shows a success-rate by approximately 70%. One of the main causes for the missing 30% seems to be the disregard of variations in position. For improvements Behiels et al. suggested an enhancement of the model with these variations. However, the risk of getting an unspecific model may be increased by the mixture of the independent variations in shape and position.

The application of ASM in plane radiography requires high performance of the algorithm. An essential criterion is the optimal shape model for covering the main shape variations. This depends mainly in correct corresponding landmark-points between different objects. One may use characteristic points, which are carefully chosen by an expert, to make sure that drawn points in different shapes are identical. These points may refer to the same anatomical landmark position on the object boundary, but may be error prone and subjective. This may only work for some characteristic points and a small number of subjects. To solve this problem, Behiels et al. proposed a bootstrap procedure to construct the training set of manual segmentation. We use a prior optimized shape model, constructed in using the minimum description length method (MDL) proposed by Davies et al. [6].

The ASM segmentation method tries to find the shape and the pose parameters of the PDM generated contour, which fits the object boundary to an image in best quality. When target points have been determined for all ASM points, the pose and shape parameters of the ASM contour are updated, so that the new ASM shape fits the target points as closely as possible. This process is repeated until convergence or a limit of iterations is reached. In the standard technique target points are determined independently for each contour point and may produce occasionally outliers that pull the contour away from the true boundary position that cannot be corrected during further iterations. This may happen especially due to lack of contrast or when multiple edges are presented nearby. Behiels et al. propose a regularization term before updating as an effective method. In this work we use a profile scale space method.

The key point for successful segmentation, however, depends on the initial placement of the template. If the initial position is too far away, the correct femur contours will never be reached. Behiels et al. have mentioned an initial instantiation of the ASM contour sufficiently close to the true object boundary contour. Within this work we try to quantify the distance of the template placement from the target shape for successful segmentation.

In Section 2 we present the idea of the knowledge-based location technique and summarize the optimized ASM technique for the femur segmentation on hip plane-radiography. Experiments evaluating the accuracy of our method are presented in Section 3. Benefits and further problems are discussed followed by a conclusion.