Detection of Sub-resolution Dots in Microscopy Images Karel

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.

Outline n n Introduction Existing approaches Method evaluation Future work

Introduction n n 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

Observable Parts of a Cell n n Cytoplasm, cytoskeleton, nucleus Whole chromosomes ¨ n n 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)

Observable Parts of a Cell – Dots n n Cytoplasm, cytoskeleton, nucleus Whole chromosomes ¨ n n 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)

Fluorescence Dots n n n 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

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

Existing Approaches to Dot Detection

“Classical” Detection Foundations n Thresholding ¨ n Mathematical morphology ¨ n Otsu, unimodal, adaptive Opening, top-hat Pattern matching

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

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

Measuring Success Rate

Measuring Success Rate n 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 ¨ n Real testing data, manual GT Different people, or the same person over multiple attempts, generally annotate images differently ¨ Time consuming, expensive ¨ n Synthetic testing data, generated GT GT is accurate and undebatable (created before the images) ¨ The synthetic data must correspond to the real images ¨

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

Recent Survey by I. Smal et al. n n Compared performance of several methods 2 D data Real images ¨ Simplified synthetic images n Dots represented by Gaussian profiles ¨ n n 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): 282– 301 (2010)

Parametrization – No Size Fits All n n 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 ¨

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

Parametrization – Usability n Proposed new method for evaluating sensitivity to parameter settings: Mean squared magnitude of the F 1 gradient, ¨ Higher value = higher sensitivity ¨ 0. 097 0. 186 1. 748

Intermediate results

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

Further Work n Investigate the conceptual difference between 2 D and 3 D 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 3 D do not take any of this into account ¨ n Design a method natively working with 3 D images Most of the existing methods are natively 2 D (or n. D), and use no special approach for 3 D data ¨ Include the new method in the comparison ¨