DiffNA: Reliable Generative Medical AI with Conditional Noise and Anatomy Guidance

"Deep learning models are reliant on large labeled datasets for optimal results across image
analysis tasks across medical imaging and other domains. Considering the challenges
associated with annotating medical images (observer variations, privacy preservation, anatomical
variations, rare diseases, data diversity, etc.), generating images with generative AI (GenAI) to
augment small datasets is an attractive alternative to curating large labeled medical image
datasets.
Understanding the diagnostic quality of medical images is extremely important for effective
image-based decision making. Particularly in computed tomography (CT), image quality is
related to radiation dose. Lower dose is safer, but lowering the dose generally leads to higher
image noise and negatively affects the subsequent image-based analyses and medical
interpretation. One of the fundamental challenges to assessing image quality is the requirement
of a high-quality reference, which can only be obtained with very high dose levels. Moreover,
subjective assessments requiring manual processes do not provide realistic guidelines for
clinically relevant downstream diagnostic tasks.
In this proposal, we will use an innovative image quality and anatomy guided generative medical
AI approach leveraging a large number of natural images and a small number of medical (CT)
images. We will evaluate our approach in assessing the quality of CT images from multiple body
locations. By performing clinically relevant downstream tasks, we will further validate the
proposed approach. Our findings will dramatically advance the understanding of CT image
quality by means of natural images, leading to significantly more effective image analyses."