Early Cerebral Palsy Screening through Deep Learning

Utilising computer vision to aid in the early diagnosis of cerebral palsy in infants

Challenge

Clinicians typically use observations of infant movement to diagnose cerebral palsy. However, this is a subjective, complex and costly process for which trained assessment experts are limited. For early detection to occur more broadly, automated technological solutions with high efficacy are urgently required.

Impact

Early detection of neurodevelopmental disorders such as cerebral palsy can radically influence timely interventions that can significantly impact a person's quality of life. Automated, high efficacy solutions made possible by state-of-the-art computer vision will enable a faster, more convenient, and lower cost cerebral palsy screening process.

Research Partners

  • Cerebral Palsy Alliance
  • Perth Children's Hospital
  • National Health and Medical Research Council

In infants, signature movements indicating a risk of cerebral palsy (CP) are weak and therefore easily overwhelmed by more dominant but irrelevant gross body motions. Furthermore, such signatures are highly specific and subtle, making them normally only recognisable by medical experts. These intricacies pose major challenges for computer vision system to recognise and extract those movements from video signals.

However, A2I2’s computer vision research program is making significant progress in analysing videos of infants and babies for CP detection. This work is funded by the NHMRC and in collaboration with our medical partners including Cerebral Palsy Alliance and Perth Children’s Hospital. Our recently published method for CP early detection achieves highest accuracy ever reported for this problem.

Image (C) Baby Moves and VICS trials

Solution

We advance computer vision and AI methods for detecting early signs of cerebral palsy from consumer-grade videos. Our research enables small body motion that are signatures for CP risk to be intelligently identified and separated from other confounding movements. These faint but informative signals are commonly overlooked by existing video analysis methods. Our method first extracts robust deep-learning based representation of body movements. These features are then analysed by advanced spatio-temporal graph convolutional networks and attention mechanisms to recognise signature pattern of CPs.

We are advancing methods in order to provide a clear explanation as to why and how the detection was made, identifying the underlying connection between the cause and effect of the signature movement, and learning from large-scale unlabelled data. 

Read about our work here