HHS Public Access Author manuscript Author Manuscript
Proc SPIE Int Soc Opt Eng. Author manuscript; available in PMC 2017 June 01. Published in final edited form as: Proc SPIE Int Soc Opt Eng. 2017 February 11; 10140: . doi:10.1117/12.2255671.
FOXP3-stained image analysis for follicular lymphoma: Optimal adaptive thresholding with maximal nucleus coverage C. Senaras1,*, M. Pennell2, W. Chen3, B. Sahiner3, A. Shana'ah4, A. Louissaint5, R.P. Hasserjian5, G. Lozanski4, and M. N. Gurcan1 11Department
Author Manuscript
2Division
of Biomedical Informatics, The Ohio State University
of Biostatistics, College of Public Health, The Ohio State University
3Office
of Science and Engineering Laboratories, Center for Devices and Radiological Health, Food and Drug Administration, S 4Department
of Pathology, The Ohio State University
5Department
of Pathology Massachusetts General Hospital
Abstract
Author Manuscript
Immunohistochemical detection of FOXP3 antigen is a usable marker for detection of regulatory T lymphocytes (TR) in formalin fixed and paraffin embedded sections of different types of tumor tissue. TR plays a major role in homeostasis of normal immune systems where they prevent auto reactivity of the immune system towards the host. This beneficial effect of TR is frequently “hijacked” by malignant cells where tumor-infiltrating regulatory T cells are recruited by the malignant nuclei to inhibit the beneficial immune response of the host against the tumor cells. In the majority of human solid tumors, an increased number of tumor-infiltrating FOXP3 positive TR is associated with worse outcome. However, in follicular lymphoma (FL) the impact of the number and distribution of TR on the outcome still remains controversial.
Author Manuscript
In this study, we present a novel method to detect and enumerate nuclei from FOXP3 stained images of FL biopsies. The proposed method defines a new adaptive thresholding procedure, namely the optimal adaptive thresholding (OAT) method, which aims to minimize undersegmented and over-segmented nuclei for coarse segmentation. Next, we integrate a parameter free elliptical arc and line segment detector (ELSD) as additional information to refine segmentation results and to split most of the merged nuclei. Finally, we utilize a state-of-the-art super-pixel method, Simple Linear Iterative Clustering (SLIC) to split the rest of the merged nuclei. Our dataset consists of 13 region-of-interest images containing 769 negative and 88 positive nuclei. Three expert pathologists evaluated the method and reported sensitivity values in detecting negative and positive nuclei ranging from 83-100% and 90-95%, and precision values of 98-100% and 99-100%, respectively. The proposed solution can be used to investigate the impact of FOXP3 positive nuclei on the outcome and prognosis in FL.
*
[email protected]; phone: (614) 688-9753.
Senaras et al.
Page 2
Author Manuscript
Keywords cell nuclei detection; optimal adaptive thresholding; histopathology; FOXP3; Follicular Lymphoma
1. Description of Purpose
Author Manuscript
Although there are several studies which focus on classification of follicular lymphoma (FL) [1-3], the impact of the number and distribution of T-lymphocytes on the outcome remains controversial. Some authors describe an association of increased T lymphocytes (TR) with improved survival [4], [5] and [6]; while others show a significantly negative impact with poorer survival and higher rates of progression to higher grade tumors [7]. These wide differences in outcomes may be related to the fact that the majority of these studies used tissue microarrays (TMAs) instead of whole sections of tumors that by design do not take into consideration FL heterogeneity. Moreover, enumeration of FOXP3 positive nuclei in these studies was based on manual methods that are notorious for poor reproducibility and high reader bias even among expert pathologists [8]. These discordant results create the need for an objective, accurate and highly reproducible computer based method capable of enumerating FOXP3 positive nuclei in high-resolution images of whole sections of FL.
Author Manuscript
In this study, we propose a novel method to detect both the positive nuclei stained with FOXP3 and the negative nuclei counterstained with hematoxylin in images of FL. In order to accurately enumerate positive nuclei as a percentage of the total nuclei, one needs to know the total number of nuclei including both positive and negative nuclei. Unlike existing cell detection studies [9, 10], our method aims to detect the positive and negative nuclei separately, so their ratio can be calculated. To extract a coarse segmentation result, we define a new approach called optimal adaptive thresholding (OAT) that translates the nuclei detection problem into a thresholding problem. After a set of morphological postprocessing operations, the algorithm analyzes the initial segmentation mask and finds possible elliptical arcs and line segments to split potentially merged cells. Finally, we integrate a super-pixel method to split the rest of the merged nuclei.
2. Methodology The nuclei detection method includes two main steps: a coarse segmentation and nuclei refinement (Figure 1). The proposed method follows similar steps for negative and positive nuclei detection. One important difference lies in the generated score map, used in the coarse segmentation.
Author Manuscript
2.1 Coarse segmentation a. Score map generation—In this step, the algorithm aims to create an initial mask of candidate nuclei which will be further refined in the next steps. The algorithm starts with creating the grayscale score maps,
and
, as follows:
Proc SPIE Int Soc Opt Eng. Author manuscript; available in PMC 2017 June 01.
Senaras et al.
Page 3
Author Manuscript
(1)
(2)
Where
, , and
are the blue, green and red color channels of the transformed RGB
image(I*) respectively. Thus, the generated score maps, and and positive nuclei with lower intensity values, respectively.
, will represent negative
Author Manuscript
b. Optimal adaptive thresholding (OAT) with maximal nucleus coverage—After the score maps are generated, the method aims to find the optimal threshold value that is able to segment the nuclei with the goal of minimizing the overall under-segmentation and over-segmentation (merged nuclei) problem, i.e. obtaining maximal nucleus coverage. This is achieved by defining a new metric M(Th) for a given threshold value, Th, as follows:
(3)
Author Manuscript
where ITh is the binarization of the score map for a given threshold value Th, ‖.‖ is the set cardinality, and CC is a function which gives the number of the connected components, whose area is between w1×Aavg (average area of a nucleus) and w2× Aavg. (The selection of Aavg w1 and w2 parameters is described in experimental setup section.) Subsequently the algorithm calculates the M values for Th ∈ [0, 1] (Figure 2-c). Finally, the algorithm calculates the optimal threshold value, T, as follows:
(4)
Author Manuscript
This method allows us to find the optimum threshold value even if the intensity histogram of the score map is not suitable for threshold selection (Figure 2-b). The output of the initial thresholding is given in Figure 2-e. Finally, the algorithm obtains the coarse segmentation result after a series of post-processing steps for noise suppression. This procedure is applied to the score maps defined by equations (1) and (2) separately for coarse segmentation of negative and positive nuclei, respectively. 2.2 Nucleus segmentation refinement a. Using elementary geometric features—We would have told that the previous step results in single as well as clumped nuclei. In the nucleus refinement part, the algorithm aims to split clumped nuclei. For that purpose, the algorithm analyzes the connected Proc SPIE Int Soc Opt Eng. Author manuscript; available in PMC 2017 June 01.
Senaras et al.
Page 4
Author Manuscript
components in the initial segmentation. The algorithm starts with finding nuclei clumps, connected components whose area is larger than w2× Aavg. After that, the algorithm finds elementary geometric features (elliptical arcs and segment lines) for each clump by using the state-of-the-art parameter-free feature detector called Ellipse and Line Segment Detector (ELSD) [12]. The algorithm analyzes the detected elliptical arcs and segment lines, which can be possible nuclei boundary segments. After the validation, the algorithm estimates potential splitting regions and apply morphological operations to prune cells (Figure 3). b. Using super-pixel method—Although the geometric features are useful for most of the merged nuclei, the fuzzy nature of the nuclei boundaries becomes a major drawback for some of the nuclei clusters. In order to solve that problem, we applied Simple Linear Iterative Clustering (SLIC), one of the state-of-the-art super-pixel algorithms [13] to the rest of the merged nuclei.
Author Manuscript
3. Experimental Setup In this study, follicular lymphoma biopsies immunostained with FOXP3 from 10 patients were scanned with Hamamatsu Photonics KK at 40× magnification. For initial experiments, an expert hematopathologist selected 14 different region-of-interest images. The images were 256 × 256 pixels and had different cell densities and color properties (Figure 4). We randomly selected one of these images to estimate avearage area of a nuclueus (Aavg), w1, and w2. (Aavg is set to 550 pixels, w1 and w2 as 0.5 and 1.5, respectively in our experiments). The rest of the images are used for evaluation.
Author Manuscript
The proposed method was evaluated by three expert pathologists. During the evaluation, they manually identified any false positive nuclei (nuclei scored as positive, but which appeared negative) and false negative nuclei (nuclei scored as negative, but which stained positively). From this, we calculated precision and recall (also known as sensitivity) values for each nucleus type [14].
4. Results and Discussion Cell detection results for both positively and negatively stained cells for two example images are presented in Figure 4. The precision-recall graph in Figure 4 summarizes the evaluation of the three expert pathologists. Recall, or sensitivity, ranged from 83-100% for negative nuclei and 90-95% for positive nuclei. Precision is defined as
Author Manuscript
and ranged from 98-100% for negative nuclei and 99-100% for positive nuclei. A BlandAltman analysis (Figure 5) revealed that the pathologists exhibited good agreement in counting positive cells. Pathologists exhibited less agreement in counting negative cells and the differences increased as the number of negative cells increased. Proc SPIE Int Soc Opt Eng. Author manuscript; available in PMC 2017 June 01.
Senaras et al.
Page 5
Author Manuscript
5. Conclusions In this study, we proposed a novel approach to detect positive and negative nuclei in FOXP3 stained follicular lymphoma images. The method defines a new adaptive thresholding procedure which aims to minimize under-segmented and over-segmented nuclei. After the initial segmentation, we extract geometric features and super-pixel regions in order to split merged nuclei. The method exhibited good sensitivity and precision in detecting both negative and positive nuclei. The pathologists who assessed the method tended to disagree in the number of negative cells in an image; inter-reader variability issues such as this could potentially be alleviated by supplementing routine clinical examinations with an automated cell detection method like the one proposed in this study.
Acknowledgments Author Manuscript
This work was funded in part by Award Number R01CA134451 and U24CA199374 (PI: Gurcan) from the National Cancer Institute. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Cancer Institute, or the National Institutes of Health. GL and MG are senior authors.
References
Author Manuscript Author Manuscript
1. Fauzi MFA, Pennell M, Sahiner B, Chen W, Shana'ah A, Hemminger J, et al. Classification of follicular lymphoma: the effect of computer aid on pathologists grading. BMC medical informatics and decision making. 2015; 15:1. [PubMed: 25889846] 2. Samsi S, Lozanski G, Shanarah A, Krishanmurthy AK, Gurcan MN. Detection of follicles from IHC-stained slides of follicular lymphoma using iterative watershed. IEEE Transactions on Biomedical Engineering. 2010; 57:2609–2612. [PubMed: 20639174] 3. Belkacem-Boussaid K, Samsi S, Lozanski G, Gurcan MN. Automatic detection of follicular regions in H&E images using iterative shape index. Computerized Medical Imaging and Graphics. 2011; 35:592–602. [PubMed: 21511436] 4. Tzankov A, Meier C, Hirschmann P, Went P, Pileri SA, Dirnhofer S. Correlation of high numbers of intratumoral FOXP3+ regulatory T cells with improved survival in germinal center-like diffuse large B-cell lymphoma, follicular lymphoma and classical Hodgkin's lymphoma. haematologica. 2008; 93:193–200. [PubMed: 18223287] 5. Carreras J, Lopez-Guillermo A, Fox BC, Colomo L, Martinez A, Roncador G, et al. High numbers of tumor-infiltrating FOXP3-positive regulatory T cells are associated with improved overall survival in follicular lymphoma. Blood. 2006; 108:2957–2964. [PubMed: 16825494] 6. Wahlin BE, Aggarwal M, Montes-Moreno S, Gonzalez LF, Roncador G, Sanchez-Verde L, et al. A unifying microenvironment model in follicular lymphoma: outcome is predicted by programmed death-1–positive, regulatory, cytotoxic, and helper T cells and macrophages. Clinical Cancer Research. 2010; 16:637–650. [PubMed: 20068089] 7. Farinha P, Al-Tourah A, Gill K, Klasa R, Connors JM, Gascoyne RD. The architectural pattern of FOXP3-positive T cells in follicular lymphoma is an independent predictor of survival and histologic transformation. Blood. 2010; 115:289–295. [PubMed: 19901260] 8. Rizzardi AE, Johnson AT, Vogel RI, Pambuccian SE, Henriksen J, Skubitz AP, et al. Quantitative comparison of immunohistochemical staining measured by digital image analysis versus pathologist visual scoring. Diagnostic pathology. 2012; 7:1. [PubMed: 22217299] 9. Neuman, U., Korzynska, A., Lopez, C., Lejeune, M. Segmentation of Stained Lymphoma Tissue Section Images. In: Piętka, E., Kawa, J., editors. Information Technologies in Biomedicine. Vol. 2. Berlin, Heidelberg: Springer Berlin Heidelberg; 2010. p. 101-113. 10. Niazi MKK, Satoskar AA, Gurcan MN. An automated method for counting cytotoxic T-cells from CD8 stained images of renal biopsies. SPIE Medical Imaging. 2013:867606–867606-10. 11. Otsu N. A threshold selection method from gray-level histograms. Automatica. 1975; 11:23–27.
Proc SPIE Int Soc Opt Eng. Author manuscript; available in PMC 2017 June 01.
Senaras et al.
Page 6
Author Manuscript
12. Pătrăucean, V., Gurdjos, P., Von Gioi, RG. Computer Vision–ECCV 2012. Springer; 2012. A parameterless line segment and elliptical arc detector with enhanced ellipse fitting; p. 572-585. 13. Achanta R, Shaji A, Smith K, Lucchi A, Fua P, Süsstrunk S. SLIC superpixels compared to stateof-the-art superpixel methods. IEEE transactions on pattern analysis and machine intelligence. 2012; 34:2274–2282. [PubMed: 22641706] 14. Fawcett T. An introduction to ROC analysis. Pattern recognition letters. 2006; 27:861–874. 15. Agresti A, Coull BA. Approximate is better than “exact” for interval estimation of binomial proportions. The American Statistician. 1998; 52:119–126.
Author Manuscript Author Manuscript Author Manuscript Proc SPIE Int Soc Opt Eng. Author manuscript; available in PMC 2017 June 01.
Senaras et al.
Page 7
Author Manuscript Author Manuscript
Figure 1.
(a) Original Image (b) coarse segmentation result for negative samples (c) refinement using geometric features (d) mask of negative nuclei (e) visualization of detected negative and positive nuclei with green and red dots respectively.
Author Manuscript Author Manuscript Proc SPIE Int Soc Opt Eng. Author manuscript; available in PMC 2017 June 01.
Senaras et al.
Page 8
Author Manuscript Author Manuscript Figure 2.
Author Manuscript
(a) Original Image, (b) intensity histogram of the (IMATH) where red line shows the Otsu threshold[11], (c) the plot of M(Th) for Th ∈ [0, 1], (d) Otsu thresholding result (e) thresholding result for T.
Author Manuscript Proc SPIE Int Soc Opt Eng. Author manuscript; available in PMC 2017 June 01.
Senaras et al.
Page 9
Author Manuscript Figure 3.
(a) two merged nuclei (b) coarse segmentation result (c) detected elliptical arcs and segment lines which are represented by red and blue color, respectively (d) refinement result.
Author Manuscript Author Manuscript Author Manuscript Proc SPIE Int Soc Opt Eng. Author manuscript; available in PMC 2017 June 01.
Senaras et al.
Page 10
Author Manuscript Author Manuscript
Figure 4.
The example detection results for positive and negative nuclei represented by red and green dots, respectively.
Author Manuscript Author Manuscript Proc SPIE Int Soc Opt Eng. Author manuscript; available in PMC 2017 June 01.
Senaras et al.
Page 11
Author Manuscript Author Manuscript
Figure 5.
The precision-recall graph according to three expert pathologists: R1, R2 and R3. Error bars were calculated using the Agresti-Coull method[15].
Author Manuscript Author Manuscript Proc SPIE Int Soc Opt Eng. Author manuscript; available in PMC 2017 June 01.
Senaras et al.
Page 12
Author Manuscript Author Manuscript
Figure 6.
Bland-Altman plots of pairwise reader (R1, R2, R3) agreement in counting negative cells (top) and positive cells (bottom).
Author Manuscript Author Manuscript Proc SPIE Int Soc Opt Eng. Author manuscript; available in PMC 2017 June 01.
Proc SPIE Int Soc Opt Eng. Author manuscript; available in PMC 2017 June 01. Published in final edited form as: Proc SPIE Int Soc Opt Eng. 2017 February 11; 10140: . doi:10.1117/12.2255671.
FOXP3-stained image analysis for follicular lymphoma: Optimal adaptive thresholding with maximal nucleus coverage C. Senaras1,*, M. Pennell2, W. Chen3, B. Sahiner3, A. Shana'ah4, A. Louissaint5, R.P. Hasserjian5, G. Lozanski4, and M. N. Gurcan1 11Department
Author Manuscript
2Division
of Biomedical Informatics, The Ohio State University
of Biostatistics, College of Public Health, The Ohio State University
3Office
of Science and Engineering Laboratories, Center for Devices and Radiological Health, Food and Drug Administration, S 4Department
of Pathology, The Ohio State University
5Department
of Pathology Massachusetts General Hospital
Abstract
Author Manuscript
Immunohistochemical detection of FOXP3 antigen is a usable marker for detection of regulatory T lymphocytes (TR) in formalin fixed and paraffin embedded sections of different types of tumor tissue. TR plays a major role in homeostasis of normal immune systems where they prevent auto reactivity of the immune system towards the host. This beneficial effect of TR is frequently “hijacked” by malignant cells where tumor-infiltrating regulatory T cells are recruited by the malignant nuclei to inhibit the beneficial immune response of the host against the tumor cells. In the majority of human solid tumors, an increased number of tumor-infiltrating FOXP3 positive TR is associated with worse outcome. However, in follicular lymphoma (FL) the impact of the number and distribution of TR on the outcome still remains controversial.
Author Manuscript
In this study, we present a novel method to detect and enumerate nuclei from FOXP3 stained images of FL biopsies. The proposed method defines a new adaptive thresholding procedure, namely the optimal adaptive thresholding (OAT) method, which aims to minimize undersegmented and over-segmented nuclei for coarse segmentation. Next, we integrate a parameter free elliptical arc and line segment detector (ELSD) as additional information to refine segmentation results and to split most of the merged nuclei. Finally, we utilize a state-of-the-art super-pixel method, Simple Linear Iterative Clustering (SLIC) to split the rest of the merged nuclei. Our dataset consists of 13 region-of-interest images containing 769 negative and 88 positive nuclei. Three expert pathologists evaluated the method and reported sensitivity values in detecting negative and positive nuclei ranging from 83-100% and 90-95%, and precision values of 98-100% and 99-100%, respectively. The proposed solution can be used to investigate the impact of FOXP3 positive nuclei on the outcome and prognosis in FL.
*
[email protected]; phone: (614) 688-9753.
Senaras et al.
Page 2
Author Manuscript
Keywords cell nuclei detection; optimal adaptive thresholding; histopathology; FOXP3; Follicular Lymphoma
1. Description of Purpose
Author Manuscript
Although there are several studies which focus on classification of follicular lymphoma (FL) [1-3], the impact of the number and distribution of T-lymphocytes on the outcome remains controversial. Some authors describe an association of increased T lymphocytes (TR) with improved survival [4], [5] and [6]; while others show a significantly negative impact with poorer survival and higher rates of progression to higher grade tumors [7]. These wide differences in outcomes may be related to the fact that the majority of these studies used tissue microarrays (TMAs) instead of whole sections of tumors that by design do not take into consideration FL heterogeneity. Moreover, enumeration of FOXP3 positive nuclei in these studies was based on manual methods that are notorious for poor reproducibility and high reader bias even among expert pathologists [8]. These discordant results create the need for an objective, accurate and highly reproducible computer based method capable of enumerating FOXP3 positive nuclei in high-resolution images of whole sections of FL.
Author Manuscript
In this study, we propose a novel method to detect both the positive nuclei stained with FOXP3 and the negative nuclei counterstained with hematoxylin in images of FL. In order to accurately enumerate positive nuclei as a percentage of the total nuclei, one needs to know the total number of nuclei including both positive and negative nuclei. Unlike existing cell detection studies [9, 10], our method aims to detect the positive and negative nuclei separately, so their ratio can be calculated. To extract a coarse segmentation result, we define a new approach called optimal adaptive thresholding (OAT) that translates the nuclei detection problem into a thresholding problem. After a set of morphological postprocessing operations, the algorithm analyzes the initial segmentation mask and finds possible elliptical arcs and line segments to split potentially merged cells. Finally, we integrate a super-pixel method to split the rest of the merged nuclei.
2. Methodology The nuclei detection method includes two main steps: a coarse segmentation and nuclei refinement (Figure 1). The proposed method follows similar steps for negative and positive nuclei detection. One important difference lies in the generated score map, used in the coarse segmentation.
Author Manuscript
2.1 Coarse segmentation a. Score map generation—In this step, the algorithm aims to create an initial mask of candidate nuclei which will be further refined in the next steps. The algorithm starts with creating the grayscale score maps,
and
, as follows:
Proc SPIE Int Soc Opt Eng. Author manuscript; available in PMC 2017 June 01.
Senaras et al.
Page 3
Author Manuscript
(1)
(2)
Where
, , and
are the blue, green and red color channels of the transformed RGB
image(I*) respectively. Thus, the generated score maps, and and positive nuclei with lower intensity values, respectively.
, will represent negative
Author Manuscript
b. Optimal adaptive thresholding (OAT) with maximal nucleus coverage—After the score maps are generated, the method aims to find the optimal threshold value that is able to segment the nuclei with the goal of minimizing the overall under-segmentation and over-segmentation (merged nuclei) problem, i.e. obtaining maximal nucleus coverage. This is achieved by defining a new metric M(Th) for a given threshold value, Th, as follows:
(3)
Author Manuscript
where ITh is the binarization of the score map for a given threshold value Th, ‖.‖ is the set cardinality, and CC is a function which gives the number of the connected components, whose area is between w1×Aavg (average area of a nucleus) and w2× Aavg. (The selection of Aavg w1 and w2 parameters is described in experimental setup section.) Subsequently the algorithm calculates the M values for Th ∈ [0, 1] (Figure 2-c). Finally, the algorithm calculates the optimal threshold value, T, as follows:
(4)
Author Manuscript
This method allows us to find the optimum threshold value even if the intensity histogram of the score map is not suitable for threshold selection (Figure 2-b). The output of the initial thresholding is given in Figure 2-e. Finally, the algorithm obtains the coarse segmentation result after a series of post-processing steps for noise suppression. This procedure is applied to the score maps defined by equations (1) and (2) separately for coarse segmentation of negative and positive nuclei, respectively. 2.2 Nucleus segmentation refinement a. Using elementary geometric features—We would have told that the previous step results in single as well as clumped nuclei. In the nucleus refinement part, the algorithm aims to split clumped nuclei. For that purpose, the algorithm analyzes the connected Proc SPIE Int Soc Opt Eng. Author manuscript; available in PMC 2017 June 01.
Senaras et al.
Page 4
Author Manuscript
components in the initial segmentation. The algorithm starts with finding nuclei clumps, connected components whose area is larger than w2× Aavg. After that, the algorithm finds elementary geometric features (elliptical arcs and segment lines) for each clump by using the state-of-the-art parameter-free feature detector called Ellipse and Line Segment Detector (ELSD) [12]. The algorithm analyzes the detected elliptical arcs and segment lines, which can be possible nuclei boundary segments. After the validation, the algorithm estimates potential splitting regions and apply morphological operations to prune cells (Figure 3). b. Using super-pixel method—Although the geometric features are useful for most of the merged nuclei, the fuzzy nature of the nuclei boundaries becomes a major drawback for some of the nuclei clusters. In order to solve that problem, we applied Simple Linear Iterative Clustering (SLIC), one of the state-of-the-art super-pixel algorithms [13] to the rest of the merged nuclei.
Author Manuscript
3. Experimental Setup In this study, follicular lymphoma biopsies immunostained with FOXP3 from 10 patients were scanned with Hamamatsu Photonics KK at 40× magnification. For initial experiments, an expert hematopathologist selected 14 different region-of-interest images. The images were 256 × 256 pixels and had different cell densities and color properties (Figure 4). We randomly selected one of these images to estimate avearage area of a nuclueus (Aavg), w1, and w2. (Aavg is set to 550 pixels, w1 and w2 as 0.5 and 1.5, respectively in our experiments). The rest of the images are used for evaluation.
Author Manuscript
The proposed method was evaluated by three expert pathologists. During the evaluation, they manually identified any false positive nuclei (nuclei scored as positive, but which appeared negative) and false negative nuclei (nuclei scored as negative, but which stained positively). From this, we calculated precision and recall (also known as sensitivity) values for each nucleus type [14].
4. Results and Discussion Cell detection results for both positively and negatively stained cells for two example images are presented in Figure 4. The precision-recall graph in Figure 4 summarizes the evaluation of the three expert pathologists. Recall, or sensitivity, ranged from 83-100% for negative nuclei and 90-95% for positive nuclei. Precision is defined as
Author Manuscript
and ranged from 98-100% for negative nuclei and 99-100% for positive nuclei. A BlandAltman analysis (Figure 5) revealed that the pathologists exhibited good agreement in counting positive cells. Pathologists exhibited less agreement in counting negative cells and the differences increased as the number of negative cells increased. Proc SPIE Int Soc Opt Eng. Author manuscript; available in PMC 2017 June 01.
Senaras et al.
Page 5
Author Manuscript
5. Conclusions In this study, we proposed a novel approach to detect positive and negative nuclei in FOXP3 stained follicular lymphoma images. The method defines a new adaptive thresholding procedure which aims to minimize under-segmented and over-segmented nuclei. After the initial segmentation, we extract geometric features and super-pixel regions in order to split merged nuclei. The method exhibited good sensitivity and precision in detecting both negative and positive nuclei. The pathologists who assessed the method tended to disagree in the number of negative cells in an image; inter-reader variability issues such as this could potentially be alleviated by supplementing routine clinical examinations with an automated cell detection method like the one proposed in this study.
Acknowledgments Author Manuscript
This work was funded in part by Award Number R01CA134451 and U24CA199374 (PI: Gurcan) from the National Cancer Institute. The content is solely the responsibility of the authors and does not necessarily represent the official views of the National Cancer Institute, or the National Institutes of Health. GL and MG are senior authors.
References
Author Manuscript Author Manuscript
1. Fauzi MFA, Pennell M, Sahiner B, Chen W, Shana'ah A, Hemminger J, et al. Classification of follicular lymphoma: the effect of computer aid on pathologists grading. BMC medical informatics and decision making. 2015; 15:1. [PubMed: 25889846] 2. Samsi S, Lozanski G, Shanarah A, Krishanmurthy AK, Gurcan MN. Detection of follicles from IHC-stained slides of follicular lymphoma using iterative watershed. IEEE Transactions on Biomedical Engineering. 2010; 57:2609–2612. [PubMed: 20639174] 3. Belkacem-Boussaid K, Samsi S, Lozanski G, Gurcan MN. Automatic detection of follicular regions in H&E images using iterative shape index. Computerized Medical Imaging and Graphics. 2011; 35:592–602. [PubMed: 21511436] 4. Tzankov A, Meier C, Hirschmann P, Went P, Pileri SA, Dirnhofer S. Correlation of high numbers of intratumoral FOXP3+ regulatory T cells with improved survival in germinal center-like diffuse large B-cell lymphoma, follicular lymphoma and classical Hodgkin's lymphoma. haematologica. 2008; 93:193–200. [PubMed: 18223287] 5. Carreras J, Lopez-Guillermo A, Fox BC, Colomo L, Martinez A, Roncador G, et al. High numbers of tumor-infiltrating FOXP3-positive regulatory T cells are associated with improved overall survival in follicular lymphoma. Blood. 2006; 108:2957–2964. [PubMed: 16825494] 6. Wahlin BE, Aggarwal M, Montes-Moreno S, Gonzalez LF, Roncador G, Sanchez-Verde L, et al. A unifying microenvironment model in follicular lymphoma: outcome is predicted by programmed death-1–positive, regulatory, cytotoxic, and helper T cells and macrophages. Clinical Cancer Research. 2010; 16:637–650. [PubMed: 20068089] 7. Farinha P, Al-Tourah A, Gill K, Klasa R, Connors JM, Gascoyne RD. The architectural pattern of FOXP3-positive T cells in follicular lymphoma is an independent predictor of survival and histologic transformation. Blood. 2010; 115:289–295. [PubMed: 19901260] 8. Rizzardi AE, Johnson AT, Vogel RI, Pambuccian SE, Henriksen J, Skubitz AP, et al. Quantitative comparison of immunohistochemical staining measured by digital image analysis versus pathologist visual scoring. Diagnostic pathology. 2012; 7:1. [PubMed: 22217299] 9. Neuman, U., Korzynska, A., Lopez, C., Lejeune, M. Segmentation of Stained Lymphoma Tissue Section Images. In: Piętka, E., Kawa, J., editors. Information Technologies in Biomedicine. Vol. 2. Berlin, Heidelberg: Springer Berlin Heidelberg; 2010. p. 101-113. 10. Niazi MKK, Satoskar AA, Gurcan MN. An automated method for counting cytotoxic T-cells from CD8 stained images of renal biopsies. SPIE Medical Imaging. 2013:867606–867606-10. 11. Otsu N. A threshold selection method from gray-level histograms. Automatica. 1975; 11:23–27.
Proc SPIE Int Soc Opt Eng. Author manuscript; available in PMC 2017 June 01.
Senaras et al.
Page 6
Author Manuscript
12. Pătrăucean, V., Gurdjos, P., Von Gioi, RG. Computer Vision–ECCV 2012. Springer; 2012. A parameterless line segment and elliptical arc detector with enhanced ellipse fitting; p. 572-585. 13. Achanta R, Shaji A, Smith K, Lucchi A, Fua P, Süsstrunk S. SLIC superpixels compared to stateof-the-art superpixel methods. IEEE transactions on pattern analysis and machine intelligence. 2012; 34:2274–2282. [PubMed: 22641706] 14. Fawcett T. An introduction to ROC analysis. Pattern recognition letters. 2006; 27:861–874. 15. Agresti A, Coull BA. Approximate is better than “exact” for interval estimation of binomial proportions. The American Statistician. 1998; 52:119–126.
Author Manuscript Author Manuscript Author Manuscript Proc SPIE Int Soc Opt Eng. Author manuscript; available in PMC 2017 June 01.
Senaras et al.
Page 7
Author Manuscript Author Manuscript
Figure 1.
(a) Original Image (b) coarse segmentation result for negative samples (c) refinement using geometric features (d) mask of negative nuclei (e) visualization of detected negative and positive nuclei with green and red dots respectively.
Author Manuscript Author Manuscript Proc SPIE Int Soc Opt Eng. Author manuscript; available in PMC 2017 June 01.
Senaras et al.
Page 8
Author Manuscript Author Manuscript Figure 2.
Author Manuscript
(a) Original Image, (b) intensity histogram of the (IMATH) where red line shows the Otsu threshold[11], (c) the plot of M(Th) for Th ∈ [0, 1], (d) Otsu thresholding result (e) thresholding result for T.
Author Manuscript Proc SPIE Int Soc Opt Eng. Author manuscript; available in PMC 2017 June 01.
Senaras et al.
Page 9
Author Manuscript Figure 3.
(a) two merged nuclei (b) coarse segmentation result (c) detected elliptical arcs and segment lines which are represented by red and blue color, respectively (d) refinement result.
Author Manuscript Author Manuscript Author Manuscript Proc SPIE Int Soc Opt Eng. Author manuscript; available in PMC 2017 June 01.
Senaras et al.
Page 10
Author Manuscript Author Manuscript
Figure 4.
The example detection results for positive and negative nuclei represented by red and green dots, respectively.
Author Manuscript Author Manuscript Proc SPIE Int Soc Opt Eng. Author manuscript; available in PMC 2017 June 01.
Senaras et al.
Page 11
Author Manuscript Author Manuscript
Figure 5.
The precision-recall graph according to three expert pathologists: R1, R2 and R3. Error bars were calculated using the Agresti-Coull method[15].
Author Manuscript Author Manuscript Proc SPIE Int Soc Opt Eng. Author manuscript; available in PMC 2017 June 01.
Senaras et al.
Page 12
Author Manuscript Author Manuscript
Figure 6.
Bland-Altman plots of pairwise reader (R1, R2, R3) agreement in counting negative cells (top) and positive cells (bottom).
Author Manuscript Author Manuscript Proc SPIE Int Soc Opt Eng. Author manuscript; available in PMC 2017 June 01.
