The aim of the Data research stream is to develop new methods for extracting clinically meaningful information from large datasets.
Key people:
- Marta Vallejo
- Dorian Gouzou
- Qiang Wang (Proteus alumnus)
Understanding data quality and limitations
This strand of work uses data from Kronoscan, a sensing system that can image and sense in real-time at microscopic detail in new dimensions using some of the world's fastest detector technology. Current research is examining the data to understand the images produced by the system, what information can be extracted from the images, how the quality of the row data can be improved, and whether it is possible to classify the images as cancer/non-cancer with high accuracy.
The research is also exploring explanations and developing a framework to understand how the machine learning models make a decision to classify an image as cancer or non-cancer. The aim is to develop more explainable machine and deep learning algorithms and get a more informed outcome from classification.
New classifiers and models
This strand of work developed highly accurate models to do post-processing classification of FLIM images for cancer/non-cancer. Multiple traditional machine learning and advanced deep learning architectures were systematically compared to select the most suitable approaches for the data and the problem under consideration.
Further work is underway to use data from the inverted Kronoscan system to create FLIM images that can be directly correlated with macroscopic histology images, enabling direct data interpretation. The aim is to explore unsupervised image-to-image translation networks that can enable a successful automatic transformation for co-registration.
More information will be available soon. If you are interested in this area, contact Marta Vallejo.