The challenge
A pharmaceutical company had a disease detection classifier trained on images of tissue stained with one reagent. It needed a way to apply the same classifier to tissue images stained with different reagents.
A major difficulty was the absence of paired images that would allow one-to-one comparison between stains. The alternatives, such as building a new classifier and processing toolset for each stain reagent, were costly and time-consuming, so the company commissioned us to research and develop the most suitable approach.
Our solution
Within a very short period, we evaluated a range of current approaches on the client's histological images. Among them, we used Generative Adversarial Networks (GANs) for unpaired image-to-image translation. A GAN combines two or more deep learning models and is designed to generate images that are visually similar to its training set.
Alongside the framework that translates images into the required stain domain, we built an interactive visualisation tool to validate the quality of the generated images.
- Evaluation of several image translation approaches on histological images
- GAN-based unpaired image-to-image translation between stain domains
- Software framework to translate images into the domain the existing classifier expects
- Interactive visualisation tool to validate generated images
The outcome
The image-to-image translation framework and the visualisation tool allowed the client to significantly cut the time and cost of processing and classifying images of tissue stained with different reagents, reusing existing software components and models instead of building new ones for each stain.
