Evaluation of Methods for Cell Nuclear Structure Analysis from Microscopy Data

Abstract

Changes in cell nuclear architecture are regulated by complex biological mechanisms that associated with the altered functional properties of a cell. Quantitative analyses of structural alterations of nuclei and their compartments are important for understanding such mechanisms. In this work we present a comparison of approaches for nuclear structure classi cation, evaluated on 2D per-channel representations from a 3D microscopy imaging dataset by maximum intensity projection. Speci cally, we compare direct classi cation of pixel data from either raw intensity images or binary masks that contain only information about morphology of the object, but not intensity. We evaluate a number of widely used classi cation algorithms using 2 di erent cross-validation schemes to assess batch e ects. We compare obtained results with the previously reported baselines and discuss novel ndings.

Keywords

• cell nucleus morphology
• bioimage analysis
• image classi cation

-

Cell nuclear structure is regulated by underlying biological mechanisms related to cell di erentiation, development, and disease [ 3, 11, 12 ]. Changes in nuclear architecture are related to altered functional properties such as gene regulation and expression. Moreover, studies in mechanobiology show that external geometric constraints and mechanical forces that deform the cell nucleus a ect chromatin dynamics and gene and pathway activation [ 10 ]. Quantitative analyses of structural alterations of nuclear structures also have medical implications, for example, in detection of pathological conditions, such as cancer [ 12, 8 ]. Although a few algorithms have been proposed to analyze cell and nuclear phenotypes in 3D [ 2 ], the dimensionality of data, various image acquisition conditions, and great variability of cells in a population present numerous challenges for 3D image analysis. 2D image representations are computationally cheaper to operate on and often contain enough information to achieve a desired performance.

In this work we present a comparison of approaches for nuclear structure classi cation, evaluated on 2D per-channel maximum intensity projections from a large 3D microscopy imaging dataset. Speci cally, we compare direct classi cation of pixel data from either raw intensity images or binary masks, which contain only object morphology information, but not texture. We evaluate a number of widely used classi cation algorithms using 2 di erent cross-validation schemes to assess batch e ects. We demonstrate near-perfect classi cation performance using 2D data and compare our results with originally reported baselines [ 5 ]. 2 2.1

Methods

Dataset description

In this study we use 3D Cell Nuclear Morphology Microscopy Imaging Dataset [ 5 ], the biggest public dataset for nuclear structure classi cation. This dataset contains 3D volumetric microscopic cell images with corresponding nuclear and nucleolar binary masks. It includes images of cells in two phenotypic states that have been shown to exhibit di erent nuclear structure. Thus, it poses a binary classi cation problem that can be used for the assessment of cell nuclear and nucleolar phenotype analysis methods. Cells are labeled with 3 di erent uorophores: DAPI (4',6-diamidino-2-phenylindole), a common stain for the nuclei, brillarin antibody (anti- brillarin) and ethidium bromide (EtBr), both used for nucleoli staining. In the dataset original images are in 1; 024 1; 024 Z lattice (Z = f30; 50g). Every sub-volume is labeled as c0, c1, c2, representing the DAPI, anti- brillarin, and EtBr channels, respectively, Fig. 1. Binary masks are obtained by segmentation of the original data in c0 and c2 channels [ 5 ].

In this work we focus on images of primary human broblast cells. A part of this collection was subjected to a G0/G1 Serum Starvation Protocol used for cell cycle synchronization, has previously been shown to alter nuclear organization and to be re ected in changes in nuclear size and shape [ 9 ]. As a result, it contains 178 3D volumetric images of cells in the following phenotypic classes: (1) 64 subvolumes of proliferating broblasts (PROLIF), and (2) 112 sub-volumes of the cell cycle synchronized by the serum-starvation protocol cells (SS). These classes serve as two categories in a binary morphology classi cation setting. 2.2

Data preprocessing

Fluorescent labels are not always speci c to the object of interest and often produce noisy background (Fig. 1). In order to assess changes in the nuclear architecture, we rst apply nuclear masks provided with the dataset to all 3 channels of original microscopy data. Due to the anisotropy in original data, we then re-scale volumes in Z dimension by a factor extracted from the corresponding meta-data. Since each of 1; 024 1; 024 Z sub-volumes typically contains between 1 and 5 nuclei, we crop re-scaled volumes into smaller 256 256 57 sub-volumes, centered at the centroid of the corresponding nuclear mask and zero-pad them, when necessary. Finally, we produce 2D representation of subvolumes by a maximum intensity projection along the Z dimension (Fig. 2). As a result, we create a set of 999 256 256 images per channel. We compare classi cation algorithms from scikit-learn, a popular Python machine learning toolkit [ 7 ], including Gaussian Naive Bayes (NB), Linear Discriminant Analysis (LDA), k nearest neighbors classi er (kNN), support vector machines with linear (SVM) and Gaussian kernels (RBF), Random Forest (RF), Extremely Randomized Trees (ET), and Gradient Boosting (GBM). All classiers use default hyper-parameters. Every image is attened into a 1D feature vector. Feature preprocessing includes subtracting the mean and scaling to unit variance of the training set. We assign the label of the whole image to every cell extracted from it. In order to assess batch e ects in the intensity images and binary masks, we compare k-fold cross-validation (CV) scheme with the Leave-2Opposite-Groups-Out (L2OGO) scheme, suggested in [ 5 ]. L2OGO ensures that: (1) all masks derived from one image fall either in the training or testing set, and (2) testing set always contains masks from 2 images of di erent classes. 3

Results

First, we evaluate the performance of algorithms for broblast nuclear classi cation using only 2D morphological information, i.e. binary masks. We compute AUROC per chennel using 2 di erent CV schemes: 20 splits in L2OGO and a 10 times repeated 4-fold CV. Results in Table 1 do not show any apparent batch effects in the 2D classi cation setting in any of the channels, as performance levels L2OGO are only slightly lower compared to 4-fold CV. As expected, classi ers are not able to pick up complex morphological relationships from attened binary vectors, even when 3 channels are combined. Results are dominated by the morphometry features extracted from binary masks, as described in [ 5 ]. The best overall result with L2OGO is achieved by the Gaussian SVM (RBF) classi er in with AU ROC = 0:772 0:041, AU P R = 0:731 0:063, and F 1 = 0:682 0:060.

Next, we evaluate the performance using only 2D pixel intensity information. Results in Table 2 indicate possible batch e ects. The performance on the nuclear c0 channel does not bene t from the presence of additional information compared to only 2D masks. But nucleolar-stained channels c1 and c2 demonstrate 20% gain in performance even using more conservative L2OGO CV. However, L2OGO here leads to a large variance of the performance metric. On average, the EtBr channel (c2) seems to provide a sightly better representation of nucleolar structure comared to the anti- brillarin (c1). Almost all classi ers in both channels show results superior of those obtained with morphometric features, see Table 1. Combining all 3 channels gives the best result, demonstrating the complement nature of stains. The best overall result is achieved by the the Gaussian SVM (RBF) classi er with AU ROC = 0:990 0:029, AU P R = 0:980 0:040, and F 1 = 0:877 0:177). 4

Discussion

In order to establish baseline evaluation of simple pixel-based nuclear structure classi cation methods, we provide a comparison of a number of widely used machine learning algorithms on both binary and intensity 2D projections of 3D microscopic images. Although DAPI structure classi cation did not bene t from using the intensity information, our results indicate usefulness of intensities of nucleolar labels: anti- brillarin and EtBr. Nuclear morphometry extracted from L2OGO

Clf kNN 0:581 SVM 0:610 RBF 0:647 RF 0:630 ET 0:606 GBM 0:673 kNN 0:552 SVM 0:579 RBF 0:579 RF 0:579 ET 0:613 GBM 0:637 c0 0:059 0:771 0:077 0:726 0:058 0:814 0:040 0:868 0:054 0:864 0:046 0:919 0:030 0:755 0:053 0:671 0:053 0:766 0:053 0:823 0:047 0:816 0:045 0:844

Raw intensity images c1 c2 0:048 0:862 0:080 0:829 0:052 0:892 0:039 0:890 0:045 0:875 0:031 0:912 0:207 0:826 0:166 0:794 0:261 0:844 0:202 0:841 0:209 0:839 0:235 0:857

c0c1c2 0:041 0:865 0:039 0:059 0:896 0:043 0:0326 0:938 0:026 0:035 0:948 0:022 0:035 0:961 0:021 0:026 0:974 0:011 0:170 0:933 0:044 0:183 0:964 0:068 0:204 0:990 0:021 0:188 0:966 0:057 0:181 0:975 0:034 0:204 0:990 0:029 binary masks seems to re ect most of the relevant changes. Increased potential for batch e ects is only observed in classi cation of nucleolar structures in channels c1 and c2. Interestingly, combining 3 channels together seems to alleviate this issue and lead to near-perfect performance in L2OGO scheme.

Presented evaluation has a number of drawbacks and requires further investigation. First, we only use attened vectors of pixels, while there exist multiple methods for texture feature extraction, which may speed up the calculation. Alternatively, deep learning-based methods can be used for automatic feature learning [ 1, 6 ]. Second, we only evaluate performance on 2D maximum intensity projections of 3D images. Bigger study could further address similar issues in the original 3D space. Finally, we assume each nucleus in the same image to be representative of the phenotypic label that is provided for the whole image. This can be addressed by using methods that are robust to label noise [ 4 ].

References

1. Ching , T. , Himmelstein , D.S. , Beaulieu-Jones , B.K. , Kalinin , A.A. , Do , B.T. , Way , G.P. , Ferrero , E. , Agapow , P.M. , Zietz , M. , Ho man , M.M., et al.: Opportunities and obstacles for deep learning in biology and medicine . Journal of The Royal Society Interface 15 ( 141 ) ( 2018 )

2. Dufour , A.C. , Liu , T.Y., Ducroz , C. , Tournemenne , R. , Cummings , B. , Thibeaux , R. , Guillen , N. , Hero , A.O. , Olivo-Marin , J.C. : Signal processing challenges in quantitative 3-d cell morphology: More than meets the eye . IEEE Signal Processing Magazine 32 ( 1 ), 30 { 40 ( 2015 )

3. Jevtic , P. , Edens , L.J. , Vukovic , L.D. , Levy , D.L. : Sizing and shaping the nucleus: mechanisms and signi cance . Current opinion in cell biology 28 , 16{ 27 ( 2014 )

4. Kalinin , A.A. , Lisitsin , D.V. : Robust estimation of qualitative response regression models . In: Applied Methods of Statistical Analysis. Simulations and Statistical Inference. AMSA-2011 . pp. 303 { 309 ( 2011 )

5. Kalinin , A.A. , Allyn-Feuer , A. , Ade , A. , Fon , G.V. , Meixner , W. , Dilworth , D. , Jeffrey , R. , Higgins , G.A. , Zheng , G. , Creekmore , A. , et al.: 3D cell nuclear morphology: Microscopy imaging dataset and voxel-based morphometry classi cation results . In: IEEE Conference on Computer Vision and Pattern Recognition (CVPR) Workshops ( 2018 )

6. Kalinin , A.A. , Higgins , G.A. , Reamaroon , N. , Soroushmehr , S. , Allyn-Feuer , A. , Dinov , I.D. , Najarian , K. , Athey , B.D. : Deep learning in pharmacogenomics: From gene regulation to patient strati cation . Pharmacogenomics 19 ( 7 ), 629 { 650 ( 2018 )

7. Pedregosa , F. , Varoquaux , G. , Gramfort , A. , Michel , V. , Thirion , B. , Grisel , O. , Blondel , M. , Prettenhofer , P. , Weiss , R. , Dubourg , V. , et al.: Scikit-learn: Machine learning in python . Journal of Machine Learning Research 12 , 2825 { 2830 ( 2011 )

8. Rakhlin , A. , Shvets , A. , Iglovikov , V. , Kalinin , A.A. : Deep convolutional neural networks for breast cancer histology image analysis . In: Campilho , A. , Karray , F. , ter Haar Romeny, B. (eds.) Image Analysis and Recognition . pp. 737 { 744 . Springer International Publishing, Cham ( 2018 )

9. Seaman , L. , Meixner , W. , Snyder , J. , Rajapakse , I. : Periodicity of nuclear morphology in human broblasts . Nucleus 6 ( 5 ), 408 { 416 ( 2015 )

10. Uhler , C. , Shivashankar , G. : Regulation of genome organization and gene expression by nuclear mechanotransduction . Nature Reviews Molecular Cell Biology 18 ( 12 ), 717 ( 2017 )

11. Zheng , G. , Kalinin , A.A. , Dinov , I.D. , Meixner , W. , Zhu , S. , Wiley, J.W. : Rotational 3d mechanogenomic turing patterns of human colon caco-2 cells during di erentiation . bioRxiv 272096 ( 2018 ), https://doi.org/10.1101/272096

12. Zink , D. , Fischer , A.H. , Nickerson , J.A. : Nuclear structure in cancer cells . Nature reviews cancer 4 ( 9 ) ( 2004 )