Fracture Detection in an MRF-Based Hierarchical Bayesian Framework

Abstract

The previous chapter discussed the extraction of fracture points and fracture surfaces in the event of a major fracture where the bone fragments are well separated. In contrast, the focus of this chapter is on the detection of hairline fractures or minor fractures. Mandibular fractures are observed to possess certain distinct patterns in X-ray or CT images. In some cases, the fractures are observed to be hairline or minor in nature. By the terms hairline fracture or minor fracture we refer to those situations where the broken bone fragments are not visibly out of alignment or have incurred very little relative displacement, respectively. The presence of noise makes the detection and subsequent visualization of such types of fractures in X-ray or CT images a very challenging task. In case of a major fracture, i.e., fractures where the broken fragments are clearly displaced relative to each other, surgical intervention is almost mandatory. However, in the case of a hairline fracture or minor fracture, the decision regarding surgical intervention is less clear since the surgeon can rely on natural bone healing for fracture reduction without having to perform reconstructive surgery. In this chapter, we propose a Markov Random Field (MRF)-based hierarchical Bayesian paradigm for detection of hairline or minor fractures and generation of the reconstructed jaw (i.e., target pattern) in such cases. Here, we model the fracture as a local stochastic degradation of a hypothetical intact mandible. In the presence of noise, the detection and subsequent visualization of hairline fractures becomes a clinically challenging task. Furthermore, the decision regarding surgical intervention for this type of fracture is often unclear as a surgeon can choose to rely solely on natural bone healing without any surgical intervention. In addition to aiding in the detection and visualization of the hairline fracture, the generated target pattern depicts how a jaw with a hairline fracture would appear if allowed to heal naturally without explicit surgical intervention. The Bayesian estimation of the mode of the posterior distribution corresponds to the target pattern (i.e., reconstructed jaw), and the differences in intensity between the input data and the MAP estimate at specific pixel locations denote the occurrence and location of a hairline fracture.