Highlights

Using GANs to segment tumors from whole body MRI.

Introduction

This is a very simple paper with a cute idea. Not clear however how well it works in reality but at least it is worth mentioning. Segmentation is done as follows:

• First a GAN is trained on a wbMRI of healthy subjects (c.f. fig 1 for results of 4 different GANs)
• Second a real wbMRI with a tumor is considered. With it, they recover the latent vector which produces the synthetic image which is the most similar to that of the real image
• They subtract the synthetic from the real image. This results into a black image everywhere except where the tumor is.

Results

Results are illustrated in fig.2. They get better results than an [old] watershed method.

Also, StyleGAN2 gets very nice results as reported by the expert score in Table 1.