GAN_MED

Comparision of two GAN models by using Inception Score and Frechet Inception Distance

Used models:

1. Simple GAN model from Monai tutorial (https://colab.research.google.com/github/Project-MONAI/tutorials/blob/main/modules/mednist_GAN_tutorial.ipynb#scrollTo=vrh_zoD20yHi)
2. Progressive Growing GANs from pytorch GAN ZOO (https://github.com/facebookresearch/pytorch_GAN_zoo)

Data used: MedMNIST data (https://github.com/Project-MONAI/MONAI-extra-test-data/releases/download/0.8.1/MedNIST.tar.gz) - hand x-rays, breast MRI, abdomen CT

Image generating

Both models were trained with the same 3 datasets: hand x-rays, breast MRI, abdomen CT from MedMNIST. After training process new three sets of data were generated. Example images are shown below.

Progressive Growing GANs' model generated images

Fig. 1. Hand x-ray images generated using Progressive Growing GANs model

Fig. 2. Breast MRI images generated using Progressive Growing GANs model

Fig. 3. Abdomen CT images generated using Progressive Growing GANs model

Simple GAN model from Monai tutorial's model generated images

Fig. 4. Hand x-ray images generated using Simple GAN model from Monai tutorial

Fig. 5. Hand x-ray images generated using Simple GAN model from Monai tutorial

Fig. 6. Hand x-ray images generated using Simple GAN model from Monai tutorial

Quantitive methods of assesing the quality of images

In this project I used Inception Score and Frechet Inception Distance to test which one is more suitable for this kind of data.

Inception Score

The lower IS is the better quality of images and dataset diversity is.

Frechet Inception Distance

The higher FID is the better quality of images and dataset diversity is.

Calculation results

IS

FID

Ending conclusions

1. PGAN generates visibly more realistic images.
2. Assuming that PGAN images have higher quality and diversity, one can tell that for this kind of data FID metric is more suitable.