Improving The Classification of FAST Ultrasound Exams

During my time working in the Clemson University IBM Watson in the Watt creative inquiry program, I researched medical imaging in the context of FAST exams (ultrasound examinations used in trauma settings). The scans created during these exams are often of largely varying quality thanks to variances in body mass, body fat percentage, and sometimes user error. The quality of these scans can directly impact assessments made, making it extremely important these scans are completed to the highest quality possible. Work had previously been done in developing deep learning solutions for classifying these scans, but performance left room for improvement. By implementing a multi-head attention vision architecture as proposed in “On the Relevance of Temporal Features for Medical Ultrasound Video Recognition” (Smith et al., 2023), we were able to see increased performance on classification tasks despite the low data volume available. This work is being actively improved upon for use in hospital environments.