Work

Deep Generative Machine Learning to Create Synthetic 2D Ultrasound Images for Rare Disease Research

Year: 2023
Type: dissertation
Abstract: Research on rare diseases is often not considered by the artificial intelligence (AI) field due to data scarcity. Hemophilia is a rare genetic bleeding disorder that affects blood clotting. Hemophiliacs often suffer from bleeding into the joint space (hemarthrosis) that causes pain and ultimately damage. Ultrasound imaging is used for non-invasive hemarthrosis assessment. Early detection using AI is key to improve treatment efficacy and access to care. This thesis proposes a novel generation framework that includes domain knowledge from an external dataset, decreasing missing information in rare cases. By dividing the problem in content and context generation, our enriched synthetic ultrasound showed no significant differences from real data; improved model performance, especially at smaller sample sizes; outperformed a previously accepted model in synthetic quality; and provided a robust model for hemarthrosis detection. Our framework facilitates AI research in the study of rare diseases and the prototyping of AI-driven clinical solutions.
Language: English
Field: Medicine
Open Access status: green
OpenAlex ID: W7133007712
Cites: 187
Cited by: 1
Related to: 10