Generative Learning Applied on the IVYGAP Medical Imaging Data
This repository hosts model checkpoints generated as part of an MSc Data Science dissertation at Kingston University London. The project explored generative deep learning to synthesise histopathology images between in situ hybridisation (ISH) and haematoxylin & eosin (H&E) modalities using the Ivy Glioblastoma Atlas Project (IVYGAP) dataset (CD44 gene subset).
Background
Glioblastoma exhibits marked morphological and molecular heterogeneity. Integrating molecular and morphological information can aid research and support more precise tumour characterisation. This work investigated whether modern generative models can synthesise ISH and H&E images bidirectionally, enabling the virtual staining of missing modalities.
Models
Four generative models were trained and compared on 19,478 paired image patches:
- CycleGAN (unpaired) – achieved the best balance of fidelity and realism, with FID scores of approximately 60 (ISH→H&E) and 92 (H&E→ISH).
- CycleGAN (paired)
- Stable Diffusion v2.1 with histopathology discriminator
- ContourDiff
Each model was trained separately for both synthesis directions (ISH→H&E and H&E→ISH). Model checkpoints are provided here for reproducibility:
CycleGAN_Unpaired_ISHtoH&E.pthCycleGAN_Unpaired_H&EtoISH.pthCycleGAN_Paired_ISHtoH&E.pthCycleGAN_Paired_H&EtoISH.pthStableDiffusion_With_Medical_Discriminator_ISHtoH&E.ptStableDiffusion_With_Medical_Discriminator_HEtoISH.ptContourDiff_ISHtoH&E.ptContourDiff_H&EtoISH.pt
Usage
These PyTorch checkpoints can be loaded with the corresponding training scripts from the project repository. All files are large and intended for research purposes only. Please cite the dissertation when using these models.
Acknowledgements
This work was undertaken as part of an MSc Data Science dissertation at Kingston University London and was supervised by Prof. Dimitrios Makris. It makes use of publicly available data from the Ivy Glioblastoma Atlas Project.