Generating cell segmentation data with GANs

Supervisors: Veronika Cheplygina
Semester: Spring 2023
Tags: machine learning, data science, medical imaging

GANs have been proposed for generation of synthetic cell image data [1], or in other words data augmentation.

We want to perform a critical survey of the field of GANs as used for data augmentation and examine some alternatives. We believe that the notion that GANs are generating “new” data should be challenged; it in fact generates a variation of the data it is being fed for training not fundamentally new data (this is not to say GANs are useless, but they are not a panacea). In addition GANs are normally unstable to train and need massive amounts of training data.

Group invariant symmetries is an option of incorporating data properties into a segmentation model [2], which may be more efficient than generation of synthetic data. Hybrid appoaches have been proposed, which incorporate group invariant symmetries into a GAN model [3], as a way to deal with GANs inherent instabilities and need for plentiful data.

We would also like to implement 1) a GAN model (chosen through the survey work), generate synthetic data and then train a segmentation model, and 2) a segmentation model using group invariant features. 1) and 2) would then be compared quantitatively on inhouse cell image data.

  1. “Generative Adversarial Networks for Augmenting Training Data of Microscopic Cell Images”, Baniukiewicz Piotr, Lutton E. Josiah, Collier Sharon, Bretschneider Till, Frontiers in Computer Science, 2019

  2. “Group Equivariant Convolutional Networks”, Taco Cohen, Max Welling, Proceedings of The 33rd International Conference on Machine Learning, PMLR, 2016.

  3. “Group Equivariant Generative Adversarial Networks”, Neel Dey, Antong Chen, Soheil Ghafurian, ICLR 2021.

*This project is a MSc thesis in collaboration with Sartorius.

Sartorius is a multinational company biopharmaceutical industry. Veronika Cheplygina would be the supervisor from ITU, but the project is mainly taking place at Sartorius, supervised by Mattias Hansson,

If you are a KCS or KDS student wishing to apply, you are welcome to contact Sartorius directly