• MRQy: Quality assessment for MRI

  Computational Imaging for Precision Medicine: A quest for generalizable AI models  (Satish Viswanath)

A key step in the pipeline of medical image analysis and computing algorithms is evaluating variations in the appearance of imaging data, especially in large multi-institutional cohorts such as TCIA. These image artifacts occur in the form of noise, motion, inhomogeneity, ringing, motion, and aliasing in radiographic scans. While these issues can impede diagnostic interpretation, they even more significantly affect the training of machine learning and computational analysis algorithms. Currently, the standard approach for assessing medical image quality is via manual inspection, a process that is both laborious as well as subjective. We have developed MRQy, an open-source tool to assess medical image quality and determine diagnostic/computational suitability both for individual datasets as well as cohorts; for MR images acquired of any body region. MRQy leverages a Python backend which employs a combination of modality-specific image features and quality measures (e.g. signal-to-noise ratio, coefficient of variation, contrast-to-noise ratio). These results can subsequently be interrogated in a specialized HTML5 based front-end, allowing for real-time filtering, visualization, and cohort creation. We will present the use of MRQy to evaluate multi-institutional cohorts in TCIA to identify the presence of specific image artifacts, as well as site- or equipment-specific variations (i.e. batch effects); which are essential to correct prior to further downstream analysis of radiographic imagesDeveloping artificial intelligence (AI) schemes to assist the clinician towards enabling precision medicine requires “unlocking” embedded information captured by different data modalities, in an intuitive and generalizable fashion. The research in my group focuses on developing novel computational imaging features (termed “radiomic” features) together with histology or molecular data for disease characterization and treatment response evaluation in vivo. We have also developed tools and approaches towards enabling these AI models to be repeatable across imaging parameters as well as reproducible across site or scanner variations. Specific problems being addressed by us include: (a) predicting response to treatment to identify optimal therapeutic pathways, as well as (b) evaluating therapeutic response to guide follow-up procedures; in the context of clinical applications in colorectal cancers and digestive diseases.

• TCIA Update