Weigert group – Bioimage Analysis and Computational Microscopy
My group focuses on image-based machine-learning (AI) approaches to extract quantitative biological information from microscopy images. For this we develop new computer vision methods of that are robust and problem-adapted to the biological questions at hand.
Concretely we are interested in
Object detection & segmentation: How to delineate biological objects (cell, nuclei) in large 2D and 3D images?
Computational Microscopy: How to augment microscopes with problem adapted computer vision methods?
Self–supervised learning: How to extract knowledge from microscopy images and time series without (or with only few) manual annotations
Instance segmentation and classification of nuclei is an impor- tant task in computational pathology. We show that StarDist, a deep learning nuclei segmentation method that we originally developed for fluorescence microscopy, can be extended and successfully applied to histopathology images. This is substantiated by winning the Colon Nuclei Identification and Counting (CoNIC) challenge 2022.
Self-supervised dense representation learning for live-cell microscopy with time arrow prediction
We developed a self-supervised pre-training method based on time arrow prediction that learns dense image representations from raw, unlabeled live-cell microscopy videos. Our approach: we task a neural network to predict the temporal order of shuffled frames based on learned image features. Background/non-informative regions should be time-symmetric, processes such as cell divisions/cell death are inherently time-asymmetric. The image shows attribution maps from trained TAP network that indeed highlight very localized regions containing predominantly cell divisions. For a full video see: https://www.youtube.com/watch?v=_eCoZuPW598
3D Reconstruction of cell organelles in beta cells
In collaboration with groups in Dresden and Berlin, we created segmentation methods to extract the shape and location of cell organelles within pancreatic beta cells from electron microscopy images. In the rendered volume, one can find insulin producing secretory granules (orange), mitochondria (blue), golgi apartus (green), and microtubules (brown).
Spotiflow: accurate and efficient spot detection for imaging-based spatial transcriptomics with stereographic flow regression
The identification of spot-like structures in large and noisy microscopy images is an important task in many life science techniques such as imaging-based spatial transcriptomics (iST). In this joint project with the La Manno lab we developed Spotiflow, a deep-learning method that solves the spot-detection problem via deep multiscale stereographic flow regression. Spotiflow is robust to different noise conditions and generalizes across different chemistries while being up to an order-of-magnitude more time and memory efficient than commonly used methods.
