Cross-Modality Vertebrae Localization and Labeling Using Learning-Based Approaches - Spinal Imaging and Image Analysis - page 302

negatives are used at every cascade level. The training algorithm is “ biased ” to

positives, such that each positive have to be correctly classi ed but the negatives are

allowed to be misclassi ed negatives i.e., False Positives, will

be further trained in the following cascades. At run-time, while positives will go
through all cascades, most negatives can be rejected in the

ed. These misclassi

first several cascades and
do no need further evaluation. In this way, the run-time speed can be dramatically
increased. In our study, we use Adaboost [ 23 ] as the basic classi er in the cascade

framework. The output of the learned classi er AðFð x ÞÞ indicates the existence

likelihood of a landmark at x.
Our learning-based framework is general to detect different anatomical structures
in different imaging modalities due to two reasons. (1) The extended Haar wavelet
generates thousands of features. In this large feature pool, there are always some
features that are distinctive to speci

c anatomies. (2) The cascade learning
framework is able to select the most distinctive features for a speci c anatomy in a

speci c imaging modality.

In spine detection, a straightforward way to use this general detection framework
is to train detectors for each vertebrae independently. However, these trained
detectors might be confused by the similar appearances of neighboring vertebrae,
particularly in the presence of diseases or imaging artifacts. An alternative way is to
train a general detector to all vertebrae. However, due to the large shape and
appearance variability across different vertebrae, e.g., the shape and size of cervical
vertebrae are very different from lumbar vertebrae, it is very different to capture the
common characteristics of all vertebrae with one detector/classi er. By observing

these two limitations, we design a hierarchical learning scheme, which essentially
categorize vertebrae and discs into different groups and use different training

strategies based on their different characteristics. Speci cally, our learning scheme

consists of three layers, anchor vertebrae, bundle vertebrae and inter-vertebral discs.

4.2 Anchor Vertebrae

Anchor vertebrae (red ones in Fig. 2 a) are vertebrae with distinctive characteristics.
They are usually the vertebrae located at the extremes of vertebral column (e.g., C2,
S1) or at the transition border of different spine sections (e.g., C7, L1). In radiology
practices, anchor vertebrae usually provide critical evidences for labels of other
vertebrae. To leverage the distinctive characteristics of anchor vertebrae, we build
anchor vertebrae detectors as the first layer of our hierarchical learning scheme. The

learning scheme is designed to achieve two goals. First, since anchor vertebrae have
distinctive characteristics and can be identi

ed exclusively, the detectors of anchor
vertebrae should be very discriminative and only have high responses around the
speci c vertebrae centers. Second, as anchor vertebrae will be used to derive the

labels of other vertebrae, the detection of anchor vertebrae should be highly robust.
To achieve the first goal, we train anchor vertebra detectors in a very discrim-

inative way. Speci cally, we only select voxels close to the speci c anchor vertebra