Home Blog Projects Newsletter
← Back to Projects
Generative AI · Medical ImagingMay 2024 – Aug 2024 · MSc Dissertation

Fluorescence Microscopy: Generating Synthetic Images with Generative AI

Using a Pix2Pix conditional GAN to synthesise realistic fluorescence microscopy images — tackling the chronic scarcity of medical imaging data by generating high-quality training images from two distinct input types.

S

SituationThe context

Medical imaging research is bottlenecked by data scarcity. Fluorescence microscopy datasets are expensive to produce and are often locked behind confidentiality constraints, leaving deep-learning models under-trained and hard to validate. Building on the Berkeley Single Cell Computational Microscopy (BSCCM) dataset, this project asked whether generative AI could responsibly expand these datasets without compromising patient confidentiality.

T

TaskThe objective

Design and evaluate a generative pipeline that produces high-quality synthetic fluorescence images from two different input types, and validate whether a newly published white-blood-cell microscopy dataset is reliable enough for supervised learning.

A

ActionWhat I built

  • Built an image-to-image translation pipeline around the Pix2Pix conditional GAN architecture for generating fluorescence microscopy images.
  • Approach 1 — Brightfield → Fluorescence: fed brightfield microscopy images into the model to generate the corresponding fluorescence output.
  • Approach 2 — Masked-channel reconstruction: designed a novel input where 5 of the 6 fluorescence channels are randomly masked, training the model to reconstruct all 6 channels — forcing it to learn cross-channel relationships.
  • Ran rigorous training and validation cycles to benchmark image quality across both input strategies.
  • Used the generated images to augment real datasets and to validate the reliability of a newly published white-blood-cell microscopy dataset with a supervised model.
R

ResultThe outcome

  • Successfully generated high-quality synthetic fluorescence microscopy images with both the brightfield and the novel masked-channel input approaches.
  • Demonstrated that Pix2Pix can meaningfully expand medical imaging datasets — improving both diversity and volume without exposing confidential source data.
  • Confirmed the reliability of the newly published white-blood-cell dataset for supervised learning tasks, giving downstream researchers a validated resource.
6
Fluorescence channels reconstructed
2
Distinct input strategies evaluated
BSCCM
Benchmark dataset used

Tech Stack

PythonPix2PixConditional GANTensorFlow / KerasDeep LearningFluorescence MicroscopyBSCCM DatasetJupyter Notebook