Traditionally, light microscopy and histology data has been utilized as single images; in her research, Kaisa Liimatainen uses these image types in 3D context. Focus stacks and 3D histology bring new potential to biomedical analysis.
Focus stacks in machine learning -based cell detection
Cell shape affects how cells can be distinguished from focus stacks, imaged by capturing same location with different focal points. Some of the cells are easy to detect with simple methods from unfocused images, while others are undistinguishable if the image is out of focus. In development of the cell detection algorithm, the aim was to pre-train the model with one cell line and adapt it to other cell lines in unsupervised manner.
“I wanted to develop a machine learning algorithm that is fast to train, and also reduces the need for manual labor in training data creation”, Liimatainen explains.
Inference of unseen cell lines is used to generate pseudo-labels, which replace manually annotated training targets. This unsupervised domain adaptation algorithm enables cell detection even from densely growing cell cultures.
3D-histology unfolds in virtual reality
In this study, mouse prostates with tumors were visualized. As there often is other use for the serial sections, only part of them were imaged, forming so called sparse image stacks. Virtual reality program, designed for 3D histology visualization, was developed with Unreal Engine. The benefits of virtual reality in 3D histology visualization were investigated. 3D modeling of structures from sparse image stacks was challenging, and several existing methods were tested.
“I ended up implementing 3D modeling algorithm designed for sparse image stacks, with which the models can be generated at runtime directly from image stacks. When specialized modeling software is not needed, it is easy to visualize new samples”, Liimatainen clarifies.
Data has been utilized in various ways, and in VR one can interactively study e.g. textured 3D models, aligned and cropped histological sections, quantitative features and virtual stainings.
“In virtual reality, we can easily observe trends in tumor locations in the prostate, and from the sections inside the 3D models we can see tumor’s composition to its full extent”, Liimatainen describes.
Vesilahti resident Kaisa Liimatainen is currently working with 3D microrheology in Tampere University.
Public defence on Thursday 31 October
The doctoral dissertation of M.Sc. (Tech) Kaisa Liimatainen in the field of biomedical engineering titled Biomedical Image Stack Analysis with Machine Learning and Virtual Reality will be publicly examined at the Faculty of Medicine and Health Technology at Tampere University at 12:00 on Thursday, October 31, 2024, at Kauppi campus, Arvo building, Jarmo Visakorpi auditorium (Arvo Ylpön katu 34). The Opponent will be Associate Professor Mireia Crispin-Ortuzar from Cambridge University. The Custos will be Assistant Professor Pekka Ruusuvuori from Tampere University.