Machine Learning to detect MRI Biomarkers .

This study was initiated with a donation in Memory of Hannah, a loving, special Cavalier King Charles Spaniels (CKCS). Although Hannah was a healthy dog all her 16 years of life, her owner Elaine Hasty wanted to help the welfare of the CKCS breed that were in pain.

The investigation was performed by Dr Michaela Spiteri and Dr Kevin Wells (Centre for Vision, Speech & Signal Processing or CVSS) and Dr Susan (Penny) Knowler and Professor Clare Rusbridge at the School of Veterinary Medicine, University of Surrey and received the best imaging award at the European Veterinary Conference,  Helsinki in 2017.

Aim of Study: Extract imaging markers from MRI in relation to Chiari-like Malformation associated pain (CM pain) and syringomyelia (SM) in adult CKCS dogs and compare:-

  • CM pain class to asymptomatic CM controls
    • Symptomatic SM class to controls.

Materials and Methods : The dogs were diagnosed based on clinical signs
and MRI. A midline sagittal MRI of the head and neck of a CKCS from the control group was chosen as a reference. The midline sagittal MR images of 77 dogs were mapped onto the reference MRI using DEMONS (non-linear) image registration, producing a 2D deformation map for each case.For each pixel, direction and magnitude of the mapping deformation were computed.  Potential biomarkers were identified amongst these descriptors using a machine learning approach consisting of a feature selection algorithm, to identify candidate markers of CM pain or SM, and a kernelised Support Vector Machine classifier, to analyse the ability of these to successfully separate controls and clinical cases.

Results: The analysis identified 5 markers for CM pain (in the regions of the nasopharynx, soft palate, caudal nucleus, hypothalamus and 4th ventricle) and 5 markers for SM (in the regions of soft and hard palate interface x 2, soft palate, trochlear nucleus, and corpus callosum). The kernelised Support Vector Machine classifier found an area under the curve (AUC) of 81.51 for CM pain and 86.10 for SM

Conclusion:  

A machine learning approach was taken to identify MRI biomarkers which can be used to develop an objective tool for diagnosis of CM SM and is hypothesis generating for future studies on the morphometrics and genetics of this condition.