Session 14: Towards Long Term Impact of DL Models in Medical Imaging

Presentation Type

Invited

Student

Yes

Track

Tools

Abstract

The current literature suggests expansion in the research area that combines Generative Adversarial Networks (GANs) and Transfer learning (TL). Generalizability and Scalability are two important attributes to evaluate DL models to assess their long-term impact. In this research we analyze if linear combination (Data augmented TL) of these two techniques is more generalizable and scalable, or TL enabled GANs approach has better long-term impact. First, we implement the Data augment TL approach by employing DCGAN to generate synthetic chest Xray images and to pre-train a VGG-16 model. Next, we implement the TL enabled GAN method by initially training WGAN using chest and abdomen images. We then retrain the WGAN using colon Xray images to classify between normal (benign) and cancer (malignant) polyps

Start Date

2-8-2022 3:30 PM

End Date

2-8-2022 4:25 PM

This document is currently not available here.

Share

COinS
 
Feb 8th, 3:30 PM Feb 8th, 4:25 PM

Session 14: Towards Long Term Impact of DL Models in Medical Imaging

Dakota Room 250 A/C

The current literature suggests expansion in the research area that combines Generative Adversarial Networks (GANs) and Transfer learning (TL). Generalizability and Scalability are two important attributes to evaluate DL models to assess their long-term impact. In this research we analyze if linear combination (Data augmented TL) of these two techniques is more generalizable and scalable, or TL enabled GANs approach has better long-term impact. First, we implement the Data augment TL approach by employing DCGAN to generate synthetic chest Xray images and to pre-train a VGG-16 model. Next, we implement the TL enabled GAN method by initially training WGAN using chest and abdomen images. We then retrain the WGAN using colon Xray images to classify between normal (benign) and cancer (malignant) polyps