Automatic segmentation of anatomic structures in knee and lumbar spine MRI

A collaboration between Swiss AI startup Balzano and the leading MSKspecialized clinics in Switzerland, resulted in a set of algorithms that interpret MRI for classification and quantification.

Several peer-reviewed studies on these algorithms by University of Zurich have confirmed them to perform on par with the world’s leading orthopedic radiologists. The most recent AI additions from this initiative now focus on automated segmentation.

Knee cartilage
The process of quantifying anatomic structures in MRI is inefficient and costly, which is why it is performed only rarely.

Also, the inter- and intra-observer variability is rather high. In the end it is a textual description in a radiology report summarizing the degree of degeneration of knee cartilage, meniscus and bones as basis for the evaluation and observation of, for example, osteoarthritis. With only a few words about the state of knee cartilage, meniscus and bones, there is a lot of room for interpretation for other clinicians reading it.

Furthermore, in follow-up comparisons it is again time-consuming to compare two studies to see how a degeneration of an anatomic structure progressed.

The AI-based automatic segmentation of anatomic structures in musculoskeletal orthopedics developed in the ScanDiags suite of AI algorithms is a decision support tool, providing valuable quantitative information for the radiologist. The quantitative measures, including a visualization of the automatically segmented area, helps advance the standardization in reporting.

Lumbar spine
The checklist for reading lumbar spine MRI studies is rather long. Going through every vertebrae body and segmenting it in all three planes is very repetitive and exhausting when performed in high volumes.

Despite the individual assessments being less complex than other pathologies – e.g. whether a disc protrusion is present or not, or whether a vertebrae body is wedge deformed and needs treatment – this work is still repetitive and tiring and therefore becomes error prone.

With the AI-based automatic segmentation of the lumbar vertebrae bodies in the sagittal plane, a number of important measures can be extracted, which are of high value for the radiologist.

Take wedge deformity as an example. It is important to know whether a vertebral body is normal or whether it is deformed. In the latter case, the shape of the deformity needs to be assessed. An important measure is the height of the anterior and posterior plate. The height of the posterior plate is relevant for the stability of the spine. Having quantitative information about those heights at hand as a result of the automatic segmentation of the vertebrae bodies, a radiologist can come up with a more precise statement compared to a simple grading in just three stages (“mild”, “moderately” or “severely deformed”).

With the geometric information about all the vertebrae bodies, the automatic segmentation of other anatomic structures such as the thecal sac or the neuroforamina is currently being further expanded by the Balzano engineering team:

  • Thecal sac segmentation in axial plane, segmentation of nerve roots for spinal canal stenosis evaluation. Assessment of thecal sac shape for focal nerve root compression identification.
  • Neuroforamina segmentation in sagittal plane. Automatically segment the area of the neuroforamina in the middle of the pedicle, identify nerve root and fluid, derive nerve root compression.
  • Automatic segmentation of intervertebral disc and neuroforamina in axial shape and evaluate for focal nerve root compression

The ScanDiags solution
The ScanDiags service (www.scandiags. com) provides these algorithms as a regular DICOM node. It receives studies from PACS or directly from MRTs and returns its findings to any DICOM destination in a configurable format. The solution is under development, with various test and validation installations in Europe, the US and South America.