Automatic generation of artificial images of leukocytes and leukemic cells using generative adversarial networks (syntheticcellgan)

Table of Contents

Title: Automatic generation of artificial images of leukocytes and leukemic cells using generative adversarial networks (syntheticcellgan)

Authors: Kevin Barreraa, Anna Merino, Angel Molina, José Rodellar

Published: Dec 2022

Link: https://doi.org/10.1016/j.cmpb.2022.107314

Summary (Generated by Microsoft Copilot):

Introduction:

  • The study focuses on developing SyntheticCellGAN (SCG), a system for generating artificial images of leukocytes and leukemic cells using generative adversarial networks (GANs).

Challenges:

  • Difficulty in collecting large, well-annotated image sets for training models, especially for low-prevalence diseases with subtle morphological differences.

Methods:

  • SCG uses two sequential GANs: a Wasserstein GAN for low-resolution images and an image-to-image translation GAN for high-resolution images of specific cell types.

Novelties:

  • The system generates high-resolution images of both normal leukocytes and rare cell types like atypical promyelocytes and hairy cells, which are challenging to find in blood smears.

Results:

  • The generated images were evaluated through quantitative metrics, expert morphological verification, and classification accuracy tests, showing high accuracy and realism.

Performances:

  • Classifiers trained with SCG images achieved up to 100% accuracy in identifying real cells, demonstrating the effectiveness of the synthetic images for training purposes.

Limitations:

  • The collection of real images for training took several years, and the system currently focuses on a limited number of cell types.

Discussion:

  • The SCG system shows promise for augmenting image datasets and improving automatic recognition models in clinical practice, though further work is needed to include more cell types and standardize across different laboratories.