AI-Powered Clarity

Dr. Aaron Rosado in the laboratory of Dr. Judy Gichoya (Department of Radiology and Imaging Sciences) and Dr. Anthony Law (Department of Otolaryngology) has been studying super-resolution models of laryngoscopy, i.e., endoscopic imaging of the larynx. His study concerns developing super-resolution models that can be employed in laryngoscopy to address healthcare disparities and improve patient care. Specifically, he is developing AI models to enhance videos obtained with lower-quality endoscopes in resource-constrained healthcare settings. 

Aaron sought to evaluate state-of-the-art AI super-resolution models: ESRGAN, MSRRESNET, BSRGAN, USRNET, SWINIR, and VRT with a very large video set. During the Emory Cloud Cluster Proof-of-Concept (POC) project in Summer of 2022, he successfully performed AI training and inferencing with one of the models. Established on the AWS at Emory cloud Platform by OIT, the POC cloud cluster provided him with eight NVIDIA A100 GPUs, the fastest and most robust GPU on the market at the time.

Using Python and with an on-premises workstation, Aaron first processed a series of dysphonic patient videos into a total of 30,815 single frame images. 21,517 were used for training and the rest were for testing. After uploading the images to the POC cloud cluster, the OIT team and Aaron worked together to submit a Slurm Job on the cluster that launched and completed the distributed training across the eight A100 GPUs, each with 40GB of GPU memory. He then proceeded to run inferencing with the testing dataset by launching a single virtual machine, Amazon EC2, with NVIDIA A10G Tensor Core GPU on AWS at Emory.

Aaron was very pleased with the GPU hardware and performance of both the POC Cloud cluster and the single Amazon EC2. Aaron needed access to state-of-the-art GPU resources to test and train super-resolution AI models with the most advanced architectural features. He was able to complete the AI training within his desired timeframe with the GPU resources that were not available on premises and would have been difficult for him to secure otherwise. After initial tuning of the Slurm commands to allow for distributed learning among GPUs, he found the POC cloud cluster to be very easy to use. In fact, the user experience with the POC cloud cluster was no different than the experience with a traditional on-premises HPC cluster, because in both scenarios, Slurm is the user interface that abstracts out the underlying infrastructure. Aaron was also very satisfied with the assistance provided by OIT as well as from the Amazon Web Service team.  Both Judy and Aaron believed that a shared resource like this POC cloud cluster would be invaluable for advancing AI research at Emory.

Aaron’s work was selected for a podium presentation at the 2022 Fall Voice Conference. His abstract title was “Comparison of Artificial Intelligence Video Enhancement Methods on Laryngoscopy Towards Improving Care in Resource Constrained Areas.”