Statistical Morphology and Fragment Mapping of Complex Proximal Humeral Fractures

Abstract

Background and Objectives: Proximal humerus fractures (PHFs) are common in the elderly, but the treatment results are often poor. A clear understanding of fracture morphology and distribution of cortical bone loss is important for improved surgical decision making, operative considerations, and new implant designs. The aim of this study was to develop a 3D segmentation fracture mapping technique to create a statistical description of the spatial pattern and cortical bone loss of complex PHFs. Materials and Methods: Fifty clinical computed tomography (CT) scans of complex PHFs and their contralateral intact shoulders were collected. In-house software was developed for semi-automated segmentation and fracture line detection and was combined with manual fracture reduction to the contralateral template in a commercial software. A statistical mean model of these cases was built and used to describe probability maps of the fracture lines and cortical fragments. Results: The fracture lines predominantly passed through the surgical neck and between the tuberosities and tendon insertions. The superior aspects of the tuberosities were constant fragments where comminution was less likely. Some fracture lines passed through the bicipital sulcus, but predominantly at its edges and curving around the tuberosities proximally and distally. Conclusions: A comprehensive and systematic approach was developed for processing clinical CT images of complex fractures into fracture morphology and fragment probability maps and applied on PHFs. This information creates an important basis for better understanding of fracture morphology that could be utilized in future studies for surgical training and implant design.

Keywords: comminution; computed tomography; fracture morphology; fragment; probability map; proximal humerus fracture; statistical model.

Figure 1

Overview of the study methodology. The cortical fragments of the fractured proximal humerus were segmented on the CT image (1, 2), the fragment boundaries were determined (3), and the fracture was virtually reduced (4). The contralateral intact side was mirrored (A), landmarked at key anatomical features (B), and a statistical mean shape model was calculated from the data of all cases (C). The fracture lines and fragment areas were projected onto the statistical shape model and superimposed for all cases, providing spatial statistical descriptions of the fracture morphology and fragments in the form of probability maps (5D).

Figure 2

Semi-automatic segmentation of fracture fragments. The CT image (1) was automatically segmented utilizing an adaptive thresholding approach (2), remaining connected fragments were isolated with minimal user interaction to indicate the parts to be separated (3, red points). Finally, the fragments were separated automatically based on the grayscale gradients of the CT image (4).

Figure 3

Semi-automatic detection of fracture lines on the segmented CT scans (1). For each fragment, high curvature areas were identified (2, thick brown areas) and guide markers were proposed to indicate the fracture path (3, red dots), which could be manually adapted to improve the line. Finally, the fracture lines were calculated by optimizing a cost–function based on the distance to the next free boundary and the fracture angle (4, red lines; note that some fragments were not shown here for better visualization).

Figure 4

Virtual fracture reduction aided by the intact contralateral side. The fragments of the ipsilateral fracture, indicated by the various colors, (1) were reduced manually based on the surface of the mirrored contralateral bone (A) used as guide (2B). The fracture lines of the reduced fragments (3C) were projected on the contralateral surface (4D).

Figure 5

Statistical model of the proximal humerus. Surface models of the intact contralateral bones (1) were labelled manually with anatomical landmarks (2). Data of all cases (3) were used to generate the mean shape model (4).

Figure 6

Statistical description of fracture lines and the bone/comminution zones. Fracture lines were projected on the mirrored contralateral surface (1) and mapped onto the mean shape model based on homology (2). Superimposing the results of all cases provided the statistical description of fracture morphology, indicating the probability of fracture lines for each point throughout the surface (5). Similarly, the bony areas were assessed (3) and their probability map was created, indirectly indicating comminution zones (6).

Figure 7

Fracture probability map calculated for all 50 cases. Left to right: Anterior, lateral, posterior, and medial views. The color scale indicates the number of cases having fracture lines passing through the points of the surface map.

Figure 8

Fragment probability map calculated for all 50 cases. Left to right: Anterior, lateral, posterior, and medial views. The color scale indicates the number of cases having bone available at the points of the surface map.