Segmentation of actin-stained 3D fluorescent cells with filopodial protrusions using convolutional neural networks
We present the architecture, training strategy and evaluation of a convolutional neural network (CNN) designed for the segmentation of actin-stained cells in 3D+t confocal microscopy image data. The segmentation performance of the CNN is evaluated using time-lapse sequences of lung adenocarcinoma cells with three genetically distinct variants of the tubulin adaptor protein, a key protein in the process of assembly of the cell cytoskeleton, displaying three different phenotypes in regards to the morphology of the cells and in particular, to the number and length of filopodial structures. We show that the CNN significantly outperforms a baseline method based on the minimization of the Chan-Vese model using graph cuts, and we discuss the inherent benefits of using the CNN over the baseline method.
|Keywords||Cell segmentation, Chan-Vese model, Convolutional Neural Networks, Filopodia|
|Persistent URL||dx.doi.org/10.1109/ISBI.2018.8363605, hdl.handle.net/1765/108756|
|Conference||15th IEEE International Symposium on Biomedical Imaging, ISBI 2018|
Castilla, C. (Carlos), Maška, M, Sorokin, D.V. (Dmitry V.), Meijering, E, & Ortiz-De-Solorzano, C. (2018). Segmentation of actin-stained 3D fluorescent cells with filopodial protrusions using convolutional neural networks. In Proceedings - International Symposium on Biomedical Imaging (pp. 413–417). doi:10.1109/ISBI.2018.8363605