Novel Lung CT Image Synthesis at Full Hounsfield Range with Expert Guided Visual Turing Test

EasyChair Preprint 10173

Abstract

Conventional image quality metrics are unsuitable to evaluate the realism and medical accuracy of synthetically generated CT images. We describe an approach based on the concept of Visual Touring Test that engages medical professionals to assess the generated images and provide useful feedback that can inform the generative process. We first describe our approach for synthesizing large numbers of novel and diverse CT images across the full Hounsfield range using a very small annotated dataset of around thirty patients and a large non-annotated dataset with high resolution medical images. Using an anatomy exploration interface we can generate CT images with anatomies that were non-existent within either of the datasets, without compromising accuracy and quality. Our approach works for all Hounsfield windows with minimal depreciation in anatomical plausibility. We then describe our Visual Turing Test methodology in detail and show results we have obtained.

Keyphrases: GAN, Visual Touring Test, Visual Turing Test, data augmentation, deep learning

@booklet{EasyChair:10173,
  author    = {Arjun Krishna and Shanmukha Yenneti and Ge Wang and Klaus Mueller},
  title     = {Novel Lung CT Image Synthesis at Full Hounsfield Range with Expert Guided Visual Turing Test},
  howpublished = {EasyChair Preprint 10173},
  year      = {EasyChair, 2023}}