Detection of Sub-resolution Dots in Microscopy Images

Transcription

1 Detection of Sub-resolution Dots in Microscopy Images Karel Štěpka, 2012 Centre for Biomedical Image Analysis, FI MU supervisor: prof. RNDr. Michal Kozubek, Ph.D.

2 Outline Introduction Existing approaches Method evaluation Future work

3 Introduction In biology and medicine, parts of living cells can be observed using fluorescence microscopy Fluorescence in-situ hybridization (FISH) Allows us to visualize individual parts of genetic material: chromosomes or their parts Probes appear as small dots in the result Image courtesy of Wikimedia Commons

4 Observable Parts of a Cell Cytoplasm, cytoskeleton, nucleus Whole chromosomes Conditions related to the number of chromosomes (e.g. Down syndrome) Telomeres, kinetochores, centromeres Individual genes Translocations (e.g. BCR/ABL genes and their relation to certain kinds of leukemia)

5 Observable Parts of a Cell Dots Cytoplasm, cytoskeleton, nucleus Whole chromosomes Conditions related to the number of chromosomes (e.g. Down syndrome) Telomeres, kinetochores, centromeres Individual genes Translocations (e.g. BCR/ABL genes and their relation to certain kinds of leukemia)

6 Fluorescence Dots Real size as small as 10 nm In the resulting image, often 1 pixel > 60 nm Because of the diffraction limit of visible light, the magnification cannot be easily improved Due to image degradations, the sensor detects a blurred image of the dot Image of a dot has a few pixels across

7 Image Degradation Noise Multiple types, with different causes and statistical distributions: Photon shot noise (Poisson) Readout noise (Gaussian) Laser speckle noise Different methods for suppression Gaussian blurring Non-linear filters (median, non-linear diffusion) Degradation by point spread function (PSF) Present even in an ideal optical system PSF of a system can be experimentally measured

8 Existing Approaches to Dot Detection

9 Classical Detection Foundations Thresholding Otsu, unimodal, adaptive Mathematical morphology Opening, top-hat Pattern matching

10 Recent Classical-Based Methods EMax Extended maxima transform, size-based filtering Gué Top-hat, thresholding, region growing HDome HDome transformation, mean shift clustering Kozubek Gradual thresholding, size-based filtering Netten Top-hat, dot label ( sweep through all intensity levels) Raimondo Top-hat, modified unimodal thresholding, pattern matching

11 Machine Learning Approach Examine potential dot locations, classify as positive/negative Usually using a sliding sub-window May lead to excessive time consumption in 3D Training required, overtraining undesirable Training set with image patches from which the classifier learns Test set necessary to determine the quality of the classifier Neural networks AdaBoost Combination of multiple weak classifiers

12 Measuring Success Rate

13 Measuring Success Rate Comparison of the results with the ground truth (GT) We can obtain GT by manually annotating real images We can generate synthetic images together with their GT Real testing data, manual GT Different people, or the same person over multiple attempts, generally annotate images differently Time consuming, expensive Synthetic testing data, generated GT GT is accurate and undebatable (created before the images) The synthetic data must correspond to the real images

14 Measuring Success Rate Detection precision = recall = TP TP + FP TP TP + FN present not present found TP FP not found FN TN F 1 score = 2 precision recall precision + recall Distinguishing between large dots and double-dots To identify chromosomal conditions such as polysomy

15 Recent Survey by I. Smal et al. Compared performance of several methods 2D data Real images Simplified synthetic images Dots represented by Gaussian profiles Did not evaluate the influence of method parameters Good starting point Ihor Smal et al.: Quantitative Comparison of Spot Detection Methods in Fluorescence Microscopy. IEEE Transactions on Medical Imaging 29(2): (2010)

16 Parametrization No Size Fits All No method can be used on all types of images without any adjustments On the pixel level, images can be very different, even when displaying the same class of objects Noise level Base intensity Dynamic range Contrast Background (non-)uniformity Illumination artifacts Amount of objects of interest

17 Parametrization Usability Usability of a method depends on: Number of its parameters Sensitivity to parameter changes Intuitiveness of its parameters for the end user A thorough parametric study is required Curse of dimensionality Some of the methods have 4 6 free parameters

18 Parametrization Usability Proposed new method for evaluating sensitivity to parameter settings: Mean squared magnitude of the F 1 gradient, Higher value = higher sensitivity

19 Intermediate results

20 Further Work Publish the evaluation of existing methods Methods were tested on various 3D images Real, manually annotated data Simulated data with known GT Parametric study, sensitivity evaluation Prepare a set of benchmark data Cover testing of all important measurements Detection, localization, double-dots Possibly make the set publicly available through CBIA web-site

21 Further Work Investigate the conceptual difference between 2D and 3D fluorescence images Dots do not lie in the same focal plane Microscopy images exhibit strong anisotropy PSF does not correspond to simple Gaussian Per-slice processing or direct extension to 3D do not take any of this into account Design a method natively working with 3D images Most of the existing methods are natively 2D (or nd), and use no special approach for 3D data Include the new method in the comparison