Novel Methods for Microscopic Image Processing, Analysis, Classification and Compression

Transcription

1 Novel Methods for Microscopic Image Processing, Analysis, Classification and Compression Ph.D. Defense by Alexander Suhre Supervisor: Prof. A. Enis Çetin March 11, 2013

2 Outline Storage Analysis Image Acquisition Feature Extraction Compression Segmentation Classification 2 Source: Lezoray, 2002 Source:

3 Microscopic Image Compression Tissues are stained using special dyes before they are examined under the microscope, and therefore have a limited color range. Original image Original image Source: Kodak dataset Source: Lezoray, 2002 Image histogram Image histogram 3 Bin # Bin #

4 Microscopic Image Compression Mathematically, the process of doing a color transform on a given pixel can be written as follows: JPEG uses color transforms of the YCbCr flavor, where most of the energy is concentrated in the Y (luminance) component. It models the human visual system (HSV). Source: 4

5 Microscopic Image Compression We propose to use an adaptive color transform that can be incorporated into block-based compression schemes, such as JPEG and JPEG2000. The proposed scheme will create no extra overhead at the receiver. 5

6 Microscopic Image Compression We replace the standard values of the luminance values with the following color weights, leaving the chroma values as they are. For all three color channels c, we can find the new luminance coefficients as follows: 6

7 Microscopic Image Compression Predicting the current block s color content from the past blocks may result in wrong estimation in the presence of edges and corners. 7

8 Microscopic Image Compression We compare the color content of the current block s (untouched) chroma channels with the previous blocks. If the difference between the two is significant, we use the baseline transform. x c Chroma vector of current block x p Average chroma vector of previous blocks δ Threshold 8

9 Microscopic Image Compression We tested our method on two datasets: 42 natural images including the Kodak dataset and high resolution images (2k-by-1.5k). In 93% of the images, our method produced better results than standard color transforms. The average PSNR gain was up to 0.15 db. 51 microscopic images. In 100% of the images, our method produced better results than standard color transforms. The average PSNR gain was up to 0.5 db. 9

10 Microscopic Image Compression Our contributions: Efficient usage of color content in image coding, by using a DPCM-like structure that produces better results than standard color transforms and creates no coding overhead. Applicable to different coding schemes (JPEG, JPEG200, MPEG ) 10

11 Outline Storage Analysis Image Acquisition Feature Extraction Compression Segmentation Classification 11 Source: Lezoray, 2002 Source:

12 Classification of Follicular Lymphoma Images Follicular lymphoma (FL) is the second most common form of non-hodgkin lymphoma. It represents 20-25% of non-hodgkin lymphomas in the United States. The world health organization (WHO) defines three different gradings for FL Grade 1: 0-5 centroblasts (CBs) per high-power field (HPF) Grade 2: 6-15 centroblasts per HPF Grade 3: More than 15 centroblasts per HPF Grade 1 Grade 2 Grade 3 12

13 Classification of Follicular Lymphoma Images Sertel et al. have mimicked the manual approach of pathologists, i.e., identifying the number of centroblasts in the sample. Based on this, a decision on the grade of the sample can be made. Accuracy for CB detection was about 80%. 13

14 Classification of Follicular Lymphoma Images We propose to use region covariance matrices (Tuzel, 2005) for this problem. Our feature vector choice will be the following: L, a, b denote the components of the L-a-b colorspace. 14

15 Classification of Follicular Lymphoma Images H p and E p denote the projections on the H and E vectors proposed by Cosatto et al. (2008) to model Hematoxylin and Eosin (H&E) staining. Original image H-Projection E-Projection 15

16 Classification of Follicular Lymphoma Images We propose a two-stage classification scheme. Note that grade 3 (class A) has significantly different features than grades 1 and 2 (class B). Classes A and B can be distinguished by comparing the histograms of the respective test and training sets via Kullback-Leibler (KL) divergence. 16

17 Classification of Follicular Lymphoma Images For differentiating grades 1 and 2, we choose a Bayesian classifier. We use the DCT of the eigenvalue histograms from the RCM. The underlying PDF is assumed to be sparse, therefore only q coefficients are used. 17

18 Classification of Follicular Lymphoma Images Our proposed classification scheme is as follows: 18

19 Classification of Follicular Lymphoma Images For the experiments, a dataset of 90 images per grade was used. Image size was 1313-by-2137 pixels. Block size was 15-by-15 pixels. Classification accuracies are given below. Grade 1 Grade 2 Grade 3 Sparsity smoothing

20 Classification of Follicular Lymphoma Images Our contributions: Efficient modelling of FL modalities by circumventing computationally expensive segmentation. Novel sparsity smoothing classifier. 20

21 Bandwidth Selection for KDE One can interpret any type of data as samples drawn from an (usually unknown) probability density function (PDF). Source: 21 For better mathematical tractability, it is desirable to smooth histograms.

22 Bandwidth Selection for KDE There are MANY ways to smooth histograms. Example: Kernel Density Estimation (KDE) 22 Source:

23 Bandwidth Selection for KDE But how do you choose σ? Poor choice of σ may result in under- or oversmoothing. 23 Source:

24 Bandwidth Selection for KDE Mathematically, KDE can be described as follows: v i denotes the data samples and k σ denotes a kernel function (e.g. Gaussian) with a certain bandwidth σ. 24

25 Bandwidth Selection for KDE Sheather et al. (1991) were aiming to find a σ that minimizes the mean integrated square error (MISE) or its approximation, the AMISE, using a pilot estimate. 25

26 Bandwidth Selection for KDE Silverman (1986) describes how to find σ by using crossvalidation (CV). Minimizing the first two terms w.r.t. σ is equivalent to minimizing the whole expression w.r.t. σ. R can be approximated by CV: 26

27 Bandwidth Selection for KDE Jones et al. (1996) have done performance tests with several kinds of bandwidth estimators. They found that CV resulted in serious undersmoothing and that Sheather s method (1991) provided the best estimates. This is the general opinion in the statistics community. However, Loader (1999) is at odds with this broad categorization. He points out that Sheather s pilot estimate relies on a-priori assumptions that may not always be valid. 27

28 Bandwidth Selection for KDE We propose to add one term to the CV minimization that imposes sparsity in the Fourier domain on the estimate. We propose the following rule of thumb for the choice of λ: 28

29 Bandwidth Selection for KDE Köse et al. (2012) introduced filtered variation (FV), which can be applied to our method as follows: with 29

30 Bandwidth Selection for KDE We tested our method on the 15 example distributions by Marron and Wand (1992). 30 Source: Marron, Wand, 1992

31 Bandwidth Selection for KDE Our estimates consistently outperform traditional CV and Sheather s method for various data lengths measured under the following figure of merit. M s Performance measure value of Sheather s method M p Performance measure value of proposed method KL divergence gain (db) for traditional CV over Sheather N=64 N=128 N=256 N=512 N=1024 N=2048 N=4096 Average db Gain

32 Bandwidth Selection for KDE Our estimates consistently outperform traditional CV and Sheather s method for various data lengths. KL divergence gain (db) for our proposed CV with added L 1 term over Sheather N=64 N=128 N=256 N=512 N=1024 N=2048 N=4096 Average db Gain KL divergence gain for our proposed CV with added L 1 term with filter over Sheather N=64 N=128 N=256 N=512 N=1024 N=2048 N= Average db Gain

33 Bandwidth Selection for KDE in Image Thresholding Our bandwidth estimation method can be used in image thresholding. Our input here is the image s histogram. The PDFs of the upper and lower clusters are estimated. 33

34 Bandwidth Selection for KDE in Image Example segmentations: Thresholding 34

35 Bandwidth Selection for KDE in Image Thresholding We evaluated our method by comparing it with Otsu s method over a dataset of 49 images with groundtruth. Results are shown using the following figure of merit: M o Metric value of Otsu s method, M p Metric value of proposed method RNU ME RFAE EM NMHD NFDR Mean Hit (%)

36 Bandwidth Selection for KDE Our contributions: Bandwidth estimation for KDE that outperforms classical approaches (Sheather), without making significant prior assumptions about the data. Successfully used in an image thresholding algorithm that outperforms Otsu s method. 36

37 Cancer Cell Line Classification Cancer cell lines are grown in tissue culture, usually in a lab environment. They represent generations of a primary culture. 37 Source:

38 Cancer Cell Line Classification Identification of carcinoma cells has to be done at several stages of an experiment in molecular biology. Short tandem repeat (STR) analysis is being used as a standard for the authentication of human cell lines. This is a costly and non-automated process. Automated analysis will provide a fast and easy-to-use tool that can be used in laboratories to verify cell line identity. 38

39 Cancer Cell Line Classification Problem: Classify images from the following cell lines. 39 The images show a large quantity of different junctions and sharp corners.

40 Cancer Cell Line Classification We propose to use region co-difference matrices (Tuna, 2008) for feature extraction. Covariance Co-difference where 40 Co-difference computation is about 100-times faster than covariance computation for a given image.

41 Cancer Cell Line Classification The co-difference operator has some interesting properties. One can define a vector product as follows: Multiplication of a vector with a scalar Vector product of a vector with itself is a scaled version of the l 1 norm 41

42 Since our images show a lot of junctions and corners, we use dual-tree complex wavelet (DTCWT) and directional codifference features. DTCWT M θ (x,y) Cancer Cell Line Classification Source: Selesnick, 2005 Directional co-differences s α (x,y) 42

43 Cancer Cell Line Classification We use the following feature vector: We investigate the effect of normalization of the covariance and co-difference matrices. 43

44 Cancer Cell Line Classification Background subtraction is carried out by using an EM algorithm followed by morphological closing and median filtering. Original image Morphological closing and median filtering EM segmentation 44

45 Cancer Cell Line Classification Our dataset consisted of 14 different cancer cell lines recorded at 20x with 20 images per class. For classification we used an SVM with an RBF kernel and parameters C=1000 and γ=0.5 after cross-validation. Accuracies in % are given below. 45

46 Our contributions: Cancer Cell Line Classification Novel classification scheme for cancer cell line images that is considerably cheaper than traditional methods. Introduction of vector product for the co-difference operator. 46

47 Publications during this Thesis Journals: Suhre A., Kose K., Cetin A.E., Gurcan M.N., Content-adaptive color transform for image compression, Optical Engineering, 50, , 2011 Keskin F., Suhre A., Kose K., Ersahin T., Cetin-Atalay R., Cetin A.E., Image Classification of Human Carcinoma Cells Using Complex Wavelet-Based Covariance Descriptors, PLoS ONE 8(1): e doi: /journal.pone Conferences: Suhre A., Keskin F., Ersahin E., Cetin-Atalay R., Ansari R., Cetin A.E., A Multiplication-Free Framework for Signal Processing and Applications in Biomedical Image Analysis, accepted for publication in IEEE International Conference on Acoustics, Speech, and Signal Processing (ICASSP), May 2013 Suhre A., Cetin A.E., Image Histogram Thresholding Using Gaussian Kernel Density Estimation, accepted for publication in 21 st IEEE Signal Processing and Communications Applications Conference (SIU), April 2013 Keskin F., Suhre A., Ersahin T., Cetin-Atalay R., and Cetin A.E., Carcinoma cell line discrimination in microscopic images using unbalanced wavelets, in 46th Annual Conference on Information Sciences and Systems (CISS), March Suhre A., Ersahin T., Cetin-Atalay R., Cetin A.E., Microscopic image classification using sparsity in a transform domain and Bayesian learning, EUSIPCO, 2011, 19 th European Signal Processing Conference, pp , August 2011 Suhre A., Kose K., Cetin A.E., Gurcan M.N., Content-adaptive color transform for image compression, 17th IEEE International Conference on Image Processing (ICIP), 2010, pp , 26-29, Sept Suhre A., Kose K., Cetin A.E., Image compression using a histogram-based color transform, IEEE 18th Conference on Signal Processing and Communications Application (SIU 10) , May 2010 Currently under review: Suhre A., Arikan O., Cetin A.E., Bandwidth Selection for Kernel Density Estimation Using Total Variation-type Constraints, submitted to IET Letters Suhre A., Arikan O., Cetin A.E., Bandwidth Selection for Kernel Density Estimation Using Fourier Domain Constraints, submitted to IEEE Signal Processing Letters 47

48 48 Thanks and Acknowledgments