Algorithm can label parts of the spine on different types of images
A single algorithm can be used to automatically label vertebrae and discs in the spine on 3D images, regardless of the type of image used.
This streamlining can improve diagnoses of spine problems and preparation for surgery and accelerate radiological workflows, according to new research.
Magnetic resonance (MR) imaging and computed topography (CT) scans are typically used to evaluate the spine. The vertebrae and intervertebral discs need to be assigned labels to serve as an anatomical reference system for clinicians. However, such labeling is time consuming; with the number of scans growing a more automated method of labeling is needed.
Current computed-aided tools to label images have run into challenges because the same tissue can have different appearances, even within a scan. The current algorithms are also trained on one type of image and need to be retrained when being used on different images, scan parameters and even different vendors of imaging equipment. Retraining the algorithms is not practicable for clinical use or feasible for radiological software vendors.
The researchers, from Vienna, Austria, aimed to create a cross-modality labeling solution to be used in clinical practice. Their goal was to develop a fully automated computer program to localize the disc and vertebrae positions and label them that could be applied to more than one type of image, image contrasts, sequence types and scans, thus avoiding retraining.
They used entropy-optimized texture models (ETS) to fix seed points—locations against which information is tagged—in place for labeling, which enables the use of a single learning based pipeline and allows the building of computer models against texture and image variations. They combined the texture models with convolutional neural networks and trained the algorithm using ETS patches.
The sacrum, the bone at the base of the spine, was used as the starting point for labeling. The computer automatically located the sacral region, and for labeling used local disc entropy models matched to the spinal column. The model-matched position was then refined by an intensity-based template matching approach, a technique for finding small parts of an image which match a template image.
The computer program was then tested on 161 publicly available scans from different scanners.
Its detection rate of the localized seed points was 93.6 percent and reached overall labeling accuracy of 92.5 percent. No sacrum seeds were found in cervical scans (near the neck), which showed that there were no false positives. The algorithm also successfully labeled several pathologies in the scans, such as scoliosis and fractures. It performed better than support vector machine, a different kind of algorithm used for classification analysis, and better than two other convolutional neural networks that used different preprocessing strategies.
Using ETS allowed the algorithm to be applied to unseen image contrasts and CT scans not covered by training, which means the computer model can be applied when only a few images exist, say the study authors.
The study appears in the International Journal of Computer Assisted Radiology and Surgery.
“Our approach achieves high sacrum localization accuracy and shows promising labeling results. To the best of our knowledge, an algorithm able to deal with such a diverse set of MR and CT scans has not yet been presented in the literature,” the authors conclude.