Automatic Localization and Identification of Vertebrae in Arbitrary Field-of-View CT Scans

Abstract

Abstract. This paper presents a new method for automatic localization and identification of vertebrae in arbitrary field-of-view CT scans. No assumptions are made about which section of the spine is visible or to which extent. Thus, our approach is more general than previous work while being computationally efficient. Our algorithm is based on regression forests and probabilistic graphical models. The discriminative, regression part aims at roughly detecting the visible part of the spine. Accurate localization and identification of individual vertebrae is achieved through a generative model capturing spinal shape and appearance. The system is evaluated quantitatively on 200 CT scans, the largest dataset reported for this purpose. We obtain an overall median localization error of less than 6mm, with an identification rate of 81%.

Uncertainty-Driven Forest Predictors for Vertebra Localization and Segmentation

“…We suggest a cascaded pipeline, as illustrated in Figure 1. We explore two applications, namely (1) light microscopic images of zebrafish embryos, where we segment developing vertebrae, called somites, and (2) benchmark human spine CTs [4], where we localize and identify vertebrae. The most important aspect of our cascaded pipeline is the question of what to infer from a constellation model at intermediate stages of the cascade.…”

“…This is compared to standard MAP inference (as in e.g. [4,13]) and model-agnostic geodesic smoothing [7]. (2) We outperform a state-of-the-art method [5] on benchmark human spine CTs of challenging pathological cases.…”

“…State-of-the-art approaches for object localization or semantic segmentation typically employ pixel-wise forest predictors combined with MAP inference on a graphical constellation model [4,5,13] or a (super-)pixel graph [14,12], respectively. A recent trend in computer vision replaces single-level forest predictors by deep, cascaded models for feature generation, such as CNNs [8] and AutoContext Models [15].…”

“…It has been shown that better features can be generated by interleaving feature generation with MAP inference [9,11,7]. 4 In this work we take this idea a step further: Instead of interleaving feature generation with a pixel-level structured model or model-agnostic smoothing, we Shared first authors. Shared last authors.…”

“…Shared last authors. 4 Note that this is conceptually different from the classical "hierarchical" approach that, purely for the sake of pruning the search space to reduce run-time, performs feature generation and inference/model fitting multiple times on different scales.…”