[MA 2023 01] Machine Learning for Foci Detection and Quantification in Fluorescent Microscopy
Amsterdam UMC, Dept. of Medical Biology, Core Facility - Cellular Imaging
Proposed by: MSc Torec Luik, Scientific Researcher [t.t.luik@amsterdamumc.nl ]
Background
In cell biology, proteins can accumulate in a cell for a myriad of reasons, either in the nucleus or in the cytoplasm. Examples are stress granules, protein aggregates or DNA damage response foci; we shall refer to all of them as foci. We can visualize them under a fluorescent microscope as bright spots, by attaching certain organic molecules that fluoresce to these proteins. These microscope images can then be analyzed by computer vision models to extract such objects and count them.
For an in-depth example, cells contain DNA molecules in their nucleus and continuously suffer from DNA damage (> 10,000 lesions per cell per day). Damage can be caused by internal mechanisms, or external mechanisms (like radiation). DNA damage is sensed and repaired by the cell, a collective of processes called the DNA-damage response (DDR). The most dangerous type of damage are double-strand breaks (DSBs), where both strands of the DNA’s double helix are cleaved. Unrepaired DSBs can cause cell death or dormancy. Wrongly-repaired DSBs may have serious consequences for the organism as a whole, as they can lead to genetic instability and can induce cancer.
Thus, DSBs play an important role in the formation of cancer but, at the same time they can be used as a means for killing cancer cells.
Thousands of proteins used in DDR accumulate in the repair spots of DSBs, collectively termed (irradiation-induced) foci. As stated before, we can visualize this with fluorescence microscopy, allowing biologists to analyze them and set up experiments to understand mechanisms of DNA damage, DNA repair and differences in cell sensitivity to radiation.
Description of the SRP Project/Problem
The aim of the SRP is to design and validate a machine learning model for detecting and quantifying foci in images from fluorescence microscopy, to be used in the AMC’s microscopy core facility for a variety of cell biology experiments.
Our current approach consists of 2 or 3 separate workflows/steps:
1. detect and segment foci (with pretrained or finetuned StarDist or CellPose models)
2. optional: detect and segment cell/nucleus (with pretrained or finetuned StarDist or CellPose models)
3. count foci (labels) per image or per cell (by pixel overlap of 1 and 2)
StarDist and CellPose are powerful ML models trained for detecting and segmenting cells and nuclei, but they are not specifically trained for foci. Furthermore, for foci we are less interested in segmentation (and its IoU metric), as knowing where there are foci is more important than its exact shape. This could allow a new model more freedom to focus on detection accuracy instead of IoU. Finally, the neural network might be able to combine all 3 different steps in 1 network, end-to-end, where it predicts both segmentations and also a count of foci per (segmented) cell. This would require adding an extra output and metric to such a segmentation model.
Our expectation is that we can at least train a model specifically for foci with improved accuracy, but hopefully we can also simplify the workflow with an end-to-end model.
Qua data, we have 1 dataset available with images and labeled foci and we expect several more datasets to be ready and collected for this SRP.
Research questions
• Can we improve foci detection in fluorescent imaging by (re)training existing networks for nucleus and cell detection, e.g. StarDist or CellPose?
• Are there other networks that can compete with StarDist or CellPose in foci detection?
• Which model has the best performance on our datasets?
• Can we quantify the amount of foci per cell within the same network?
• Does combining the workflow in 1 network improve performance?
Expected results
A fine-tuned neural network that has improved predictive value on foci detection (over StarDist/CellPose) and (optionally) can quantify the amount of foci per cell.
A scientific article and presentation describing the methods, results, discussions and related work.
Time period (usually 7 months)
7 months
Contact
MSc Torec Luik, Scientific Researcher, Dept. of Medical Biology, Core facility - Cellular Imaging,
t.t.luik@amsterdamumc.nl
References
Stringer, C., Wang, T., Michaelos, M. et al. Cellpose: a generalist algorithm for cellular segmentation. Nat Methods 18, 100–106 (2021). https://doi.org/10.1038/s41592-020-01018-x
Schmidt, U., Weigert, M., Broaddus, C., Myers, G. (2018). Cell Detection with Star-Convex Polygons. In: MICCAI 2018. Lecture Notes in Computer Science(), vol 11071. Springer, Cham. https://doi.org/10.1007/978-3-030-00934-2_30