Carlos Arteta,
Victor Lempitsky,
Alison Noble,
and
Andrew Zisserman
Project Overview
Cell detection in microscopy images is an important step in the automation of cell based-experiments. In this project, we aim to develop a robust detector of cells (or other objects) in microscopy images that can perform across different microscopy imaging modalities. For this task, we have proposed a machine learning-based method where a cell model is learned from simple dot annotations; it requires few images for training and the learning can be done within a structured SVM framework.
Method
Cell candidate detection
Initially, we collect a representative set of extremal regions in the image to be used as candidate objects for detection. These are encoded with information of their size, color, shape and context. A key property of the extremal regions is that overlapping regions are nested.
Picking the best sub-set of regions
Each candidate region is evaluated with a model W learned from training data. The scores obtained from this evaluation are used to select the best candidate regions such that they do not overlap. Due to the nestedness property of extremal regions, the regions can be arranged in trees and the optimization problem can be solved exactly via dynamic programming.
Learning the model from dot annotations
The model W is learned from a set of training images which simply contain a dot inside each cell. This is done within a structural SVM framework by minimizing a cost function that penalizes the selection of regions in proportion to the deviation from one-to-one correspondences between regions and dots. The learning procedure will directly optimize the performance of the region selection in the tree structure.
Example Applications
Histopathology of breast cancer
The task is to detect lymphocytes in breast cancer histopathology images, where it is necessary to discriminate them from similar structures present in the image such as breast cancer cells.
Human embryonic kidney cells on fluorescence microscopy
The task is to detect stained cell nuclei in fluorescence images with low contrast, fading boundaries, cells out of focus and cell overlap.
HeLa cells on phase contrast microscopy
The task is to detect cancer cells in a culture, which can present high variability in size and shape.
Source code
The MATLAB code and a demonstration dataset are available in the software page.
Relevant Publications
Acknowledgements
This work was supported by the RCUK Centre for Doctoral Training in Healthcare Innovation (EP/G036861/1) and ERC grant VisRec no. 228180.
