NLM DIR Seminar Schedule

Abstract:

Medical imaging AI has demonstrated its potential to deliver efficiencies and improvements to healthcare through many studies in the literature and a growing number of applications seeking FDA approval. However, significant hurdles and challenges remain for effective, trustworthy, and general real-world implementation and adoption. In this talk, I will present several of our recent works for addressing some of the challenges hindering wide use. I will focus on quality control and data preprocessing which are key early components in the medical imaging AI pipeline and play a vital role in the overall AI performance. The research was mostly motivated by issues that we observed in some datasets for automating the visual evaluation in the screening of cervical cancer, oral cancer, and pulmonary diseases. Specifically, I will introduce our research in identifying low-quality relating to images (such as blur, noise), existence of unrelated data (such as non-medical, atypical images), and mislabeled images. I will also briefly describe methods to extract information (such as anatomical site, ruler) from images that would be useful in the subsequent pipeline tasks of disease detection and classification. In addition, I will touch on issues such as cross-dataset variance, data imbalance, domain shift, and catastrophic forgetting that are commonly encountered when applying AI to medical datasets.