1 / 17

Cell segmentation

Cell segmentation. Oleg Sklyar, Gregoire Pau, EMBL-EBI Cambridge gregoire.pau@ebi.ac.uk. Experimental setup. RNAi cell-array End-point assay HeLa cells Channels Actin (TRITC) Tubulin (Alexa 488) DNA (Hoechst). Cell segmentation. Challenging problem Cells superposition Noisy channels

farren
Download Presentation

Cell segmentation

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Cell segmentation Oleg Sklyar, Gregoire Pau, EMBL-EBI Cambridge gregoire.pau@ebi.ac.uk

  2. Experimental setup • RNAi cell-array • End-point assay • HeLa cells • Channels • Actin (TRITC) • Tubulin (Alexa 488) • DNA (Hoechst)

  3. Cell segmentation • Challenging problem • Cells superposition • Noisy channels • Artefacts (cracks, tubulin spots) • Knocked-down cells  little a priori information • Algorithm design issues: • Low a priori a priori knowledge, • Joint usage of channels is desirable

  4. Issues • Addressed issues • Actin protusions are sometimes outside a cell membrane • Cell membranes are sometimes underestimated (ie. flat) • Nuclei are sometimes outside a cell membrane • Pending issues • Superposed cells are often not correctly segmented (ie. Elongated) • Bi-nucleated cells and close cells are sometimes wrongly segmented (bi-nucleated are often segmented as two different cells and close different cells are sometimes segmented as unique ones)

  5. Proposed algorithm • Algorithm sketch: • Find nuclear envelopes on DNA channel • Filter “bad” nuclear envelopes • Find cells given nuclear envelopes, on Tubulin and Actin channels • Filter “bad” cells • Iterate previous steps until stabilization • Simple and easy to understand/model • Joint usage of all channels • Iterative converging algorithm

  6. Finding nuclear envelopes • Let denote by H the DNA channel • Nuclei in different condensation states • Different brightness HHH ATH

  7. Finding nuclear envelopes • Global thresholding approach • Nmask = H > t • How t can be set ? • There is no optimal t: this appoach cannot work t too large: mangled nuclei t too small: unseparated nuclei

  8. Finding nuclear envelopes • Local thresholding approach • Nmask = (H - Hm) > t' • where Hm is a local H average, Hm=HM, with window M: • M should be of size twice than the average nucleus size • OK ! Hm (H - Hm) > t' H

  9. Filtering 'bad' nuclei • Given some external rules: • Too small or too large • Too pale or too bright • Mangled because located on the borders • Too much empty space • … • Rules should be set by a biologist • But can be also automatically set (and tuned afterwards)

  10. Finding cell membranes • Cell membranes determination is done in two steps: • Compute the binary cell mask (cells / no matter) • Boundaries between cells is determined using Voronoi tesselation Step 1: Cmask Nmask Step 2: Cmask Nmask Cbound

  11. Cell mask determination • Global thresholding approach • Let be Z = A + T,a matter indication function • Cmask = (ZN > t), with a short filter N to prevent noise • N should be as small as the smallest cell detail we want to spot • Threshold t is computed such as: • Nuclei shoud be inside cells  Nmask  Cmask • Visible actin should be inside cells  (TN > v)  Cmask • Visible tubulin should be inside cells  (AN >v)  Cmask • With v, visibility threshold Actin Tubulin Actin+Tubulin Cbound

  12. Finding cell boundaries • Using Voronoi tesselation • Given a set of centers, what are the regions that contain the closest points to them ? • Voronoi graph = Dual k-means centroids graph • Using here nuclei (Nmask) as centers and Cmask as matter mask • Using an Euclidian/geodesic -mixed metric, based on Z = A + T gradient Geodesic =1e5 Euclidian =0

  13. Filtering 'bad' cells • Given some external rules: • Too small or too large • Too little or too large tubulin amount • Too little or too large actin amount • ... • Remove ‘bad’ cells • Debris • Tubulin bright spots • Cracks artefacts

  14. Iterate until stabilization • Algorithm sketch: • Find nuclear envelopes on DNA channel • Filter “bad” nuclear envelopes • Find cells given nuclear envelopes, on Tubulin and Actin channels • Filter “bad” cells • Iterate previous steps until stabilization • Iterative algorithm • Bounded (Nmask, Cmask) inclusive sequence  Convergence • Effectively carry joint information from step to step

  15. Results

  16. Results

  17. Pending Issues • Superposed cells are not correctly segmented • Require a superposed model • Much harder problem ! • Nuclei are sometimes wrongly segmented • Binucleated cells whose nuclei are too far to each other • Normal cells whose nuclei are too close together • Could be alleviated using T and A channels during nuclei segmentation

More Related