context
The limit on this research is labelled data, not ideas
Understanding what cosmic radiation does to astronauts means studying microscopy images of irradiated cells. NASA's BPS Microscopy Dataset on AWS is the public starting point, and like most scientific datasets it is smaller than the models that want to use it.
approach
Generate cells that plausibly look irradiated
We built an ML pipeline in Python and PyTorch Lightning that trains a GAN to produce synthetic radiation images, expanding what downstream models can train against.
transfer learning
Starting from MNIST, of all places
Rather than training from scratch, we initialized from an MNIST-trained CGAN. Microscopy images are not handwritten digits, and we did not expect much. The early convolutional features transferred further than they had any right to — edge and texture detectors are apparently not that picky about their subject.