Canadian PI: Dr. Abhilash Rakkunedeth Hareendranathan
Canadian Institution: University of Alberta
Indian PI: Dr. Mahesh Raveendranatha Panicker
Indian Institution: Indian Institute of Technology (IIT) Palakkad
Project Summary:
The Covid-19 pandemic exposed various vulnerabilities of healthcare systems worldwide including the inadequate reach of healthcare infrastructure and access to medical imaging facilities for large segments of society. A key challenge post covid-19 is to monitor the long-term impact and comorbidities in affected individuals. This places a huge resource burden on health systems as most of these assessments are currently subjective and require trained medical experts onsite. Currently, these examinations also require expensive imaging equipment (like CT and MRI) that are often less accessible to remote communities and disadvantaged individuals and may not be the preferred option for continuous and periodic monitoring. During the pandemic, in many cases, these communities were disproportionately affected, and represent groups in which studies on long-term impact and comorbidities are most needed.
Lung ultrasound(LUS) is more sensitive than X-ray in identifying lung involvement in post-covid-19 follow-up[1]. It is safe (radiation-free), inexpensive and ideal for studies in primary-care settings, and remote communities. With the recent introduction of low-cost point-of-care ultrasound (POCUS) probes (like Philips Lumify or Butterfly), sonographers can now carry ultrasound scanners in their pockets. Despite obvious benefits, POCUS is not universally used as the first tool in covid-19 follow-up. This is partly because ultrasound scans are hard to interpret. Also, the quality of POCUS scanners may not match that of conventional ultrasound machines currently used in hospitals.
Researchers propose AI-augmented POCUS as a cost-effective solution for studying the long-term effects of covid-19. Our key innovations are the use of AI to perform automatic interpretation of LUS and enhance the quality of POCUS by using domain adaptation (DA) techniques. A major challenge in training AI on medical data is to generate complete image-level gold-standard labels which are time-consuming and costly. To address these issues, we propose a two-fold approach 1) behavioral-driven annotation during acquisition and 2) weakly supervised learning(WSL) techniques that can learn from small amounts of data. We will also develop DA techniques to enhance the image quality of POCUS. By combining AI and POCUS, we propose a low-cost healthcare pathway that minimizes the need for experts and is well suited to reach remote communities where the need for post covid monitoring is greatest.
