A workflow-integrated brain tumor segmentation system based on fastai and MONAI

Abstract

Artificial intelligence (AI) has achieved great results in medical imaging tasks and has the potential to improve the experiences of clinicians and patients in the future, but on the way toward AI integration in medicine, there are many practical, technical, and societal challenges. In this thesis, we contribute to the development of AI integration in Helse Vest and present a brain tumor segmentation system integrated with their existing research PACS solution. We investigate to which degree integration of machine learning models is currently possible and if additional software development efforts are needed. The machine learning model used is developed with a library combining the two python-based deep learning libraries fastai and MONAI. This library is currently under development by researchers at Mohn Medical Imaging and Visualization Centre (MMIV), and we compare it with another state-of-the-art framework to quantify its potential usefulness. Additionally, we deploy it in a simple interactive web application. The thesis contains three studies that were conducted to discuss and answer our research goals. All studies used medical data from a data set coming out of the BraTS 2021 segmentation challenge, and our project is a part of MMIV's WIML project. Our achieved results open the way for future developers to continue workflow integrated machine learning in research PACS, and we see many possible directions to take future research.