Converting large medical images to embeddings to train classifier models
Project proposition from Pr. Isabelle Bloch’s team
Prat project title: Converting large medical images to embeddings to train classifier models
Introduction & Goals
Glioblastoma is one of the lethal tumors in adults. Recently, it has been discovered
that exploring tumor micro-environment (TME) will lead to better understanding of
immune invasion (the presence of antibodies to eliminate the tumor cells). For this
purpose, medical imaging field proposes high quality microscopic images in CZI format.
These images can have more than 100K pixels (also called giga-pixel images). The main
goal is to extract useful information from these large CZI images to train a classifier
model that will predict different stages (A, B, C & D) of immune invasion of brain
tumors.
Material
In collaboration with Paris Brain Institute, a data set of manually annotated images
(CZI format) was obtained. Brain tumor specialists have distinguished four classes to
describe the immune invasion of brain tumors.
To train a classifier based on this data set, one of the interesting methodologies is
to encode images in terms of embeddings. The goal is to test multiple deep learning
architectures based on a previous work by J. Yao et al. (Yao2020*). The embeddings
obtained from the different architectures can be used to train multiple classifier models.
As we have manually annotated classes, all the models can be evaluated. The best
method will be retained and used to predict the different stages of immune invasions
in the brain tumors. The code from the paper (Yao2020*) is available on github. The
student can reuse it and develop multiple other networks, after a literature review in
the first part of the project.
References
Publications:
*Yao, Jiawen, et al. “Whole Slide Images Based Cancer Survival Prediction Using
Attention Guided Deep Multiple Instance Learning Networks.” Medical Image Analysis,
vol. 65, Oct. 2020, p. 101789. DOI.org (Crossref). https://www.sciencedirect.com/science/article/abs/pii/S1361841520301535?via%3Dihub
