Hackathon: Spine AI - Detecting Scoliosis with Machine Learning

Spine AI: detecting scoliosis in ordinary pictures using machine learning and image processing

Scoliosis is a curvature of the spine that develops during growth. This abnormality occurs in 1-3% of the children and worsens in 0.3 to 0.5% of all children so eventually an operation is necessary. This operation in children is necessary because these severe scolioses can continue to increase in later life, cause lung problems and can seriously affect the quality of life. On the other hand, smaller scoliosis usually remains stable in the rest of life. Because scoliosis has the greatest risk of becoming more severe during the growth spurt, further curvature can be prevented in children with brace treatment. That is why it is very important to discover scoliosis early.

In the past, screening for scoliosis was part of the standard youth health care. The costs of this screening were estimated at about 3.5 million euros per year in the Netherlands in 2002. Because the screening for scoliosis turned out to be insufficiently effective, it has now been abolished in many countries, including the Netherlands. For Youth health care (JGZ) institutions in the Netherlands, screening is no longer mandatory and is no longer part of standard care. That is why scoliosis must now be noticed by parents. The outward characteristics of scoliosis become increasingly clear as the severity of the curvature increases. To be able to discover a mild curve, someone must be trained. Parents, however, are not trained in recognizing scoliosis. Therefore, our goal is to improve the feasibility of developing an app for the mobile phone or tablet to support both caregivers and parents in discovering mild scoliosis.

In this hackathon, we will focus on detecting scoliosis in an anonymized set of ~400 pictures taken at OLVG of patients against a uniformly-colored background. The set was anonymized by extracting only the image gradients and foreground/background color distributions; the original imagery remains at OLVG. The gradient and color features may serve as rudimentary input to machine learning algorithms. 

A set of challenges will be formulated, with varying difficulty levels. Also, there will be challenges, solving certain subtasks, that may not involve image processing but only analysis and modeling of structured data.