Immunogradient Indicators for Antitumor Response Assessment by Automated Tumor-Stroma Interface Zone Detection
2020; Elsevier BV; Volume: 190; Issue: 6 Linguagem: Inglês
10.1016/j.ajpath.2020.01.018
ISSN1525-2191
AutoresAllan G. Rasmusson, Dovile Zilenaite-Petrulaitiene, Aušrinė Nestarenkaitė, Renaldas Augulis, Aida Laurinavičienė, Valerijus Ostapenko, Tomas Poškus, Arvydas Laurinavičius,
Tópico(s)Cancer Genomics and Diagnostics
ResumoThe distribution of tumor-infiltrating lymphocytes (TILs) within the tumor microenvironment provides strong prognostic value, which is increasingly important with the arrival of new immunotherapy modalities. Both visual and image analysis–based assays are developed to assess the immune contexture of the tumors. We propose an automated method based on grid subsampling of microscopy image analysis data to extract the tumor-stroma interface zone (IZ) of controlled width. The IZ is a ranking of tissue areas by their distance to the tumor edge, which is determined by a set of explicit rules. TIL density profiles across the IZ are used to compute a set of novel immunogradient indicators that reflect TIL gradient towards the tumor. We applied this method on CD8 immunohistochemistry images of surgically excised hormone receptor–positive breast and colorectal cancers to predict overall patient survival. In both cohorts, the immunogradient indicators enabled strong and independent prognostic stratification, outperforming clinical and pathologic variables. Patients with breast cancer with low immunogradient levels had a prominent decrease in survival probability 5 years after surgery. Our study provides proof of concept that data-driven, automated, operator-independent IZ sampling enables spatial immune response measurement in the tumor-host interaction frontline for prediction of disease outcomes. The distribution of tumor-infiltrating lymphocytes (TILs) within the tumor microenvironment provides strong prognostic value, which is increasingly important with the arrival of new immunotherapy modalities. Both visual and image analysis–based assays are developed to assess the immune contexture of the tumors. We propose an automated method based on grid subsampling of microscopy image analysis data to extract the tumor-stroma interface zone (IZ) of controlled width. The IZ is a ranking of tissue areas by their distance to the tumor edge, which is determined by a set of explicit rules. TIL density profiles across the IZ are used to compute a set of novel immunogradient indicators that reflect TIL gradient towards the tumor. We applied this method on CD8 immunohistochemistry images of surgically excised hormone receptor–positive breast and colorectal cancers to predict overall patient survival. In both cohorts, the immunogradient indicators enabled strong and independent prognostic stratification, outperforming clinical and pathologic variables. Patients with breast cancer with low immunogradient levels had a prominent decrease in survival probability 5 years after surgery. Our study provides proof of concept that data-driven, automated, operator-independent IZ sampling enables spatial immune response measurement in the tumor-host interaction frontline for prediction of disease outcomes. One of the most prominent discoveries in modern genetics—analysis of the whole human genome sequence—has made it possible to determine specific genetic mutations associated with cancer. This determination has led to the definition of cancer as a cell disease caused by genetic mutations.1Stratton M.R. Campbell P.J. Futreal P.A. The cancer genome.Nature. 2009; 458: 719-724Crossref PubMed Scopus (2385) Google Scholar,2Wong K.M. Hudson T.J. McPherson J.D. Unraveling the genetics of cancer: genome sequencing and beyond.Annu Rev Genomics Hum Genet. 2011; 12: 407-430Crossref PubMed Scopus (81) Google Scholar Although genetic mechanisms explain many aspects of tumor progression, many hallmarks of cancer, including host immune and inflammatory response, angiogenesis, and metabolic disarrangements, evolve in the context of the tumor microenvironment (TME).3Hanahan D. Weinberg R.A. Hallmarks of cancer: the next generation.Cell. 2011; 144: 646-674Abstract Full Text Full Text PDF PubMed Scopus (42734) Google Scholar In particular, a major driving force of local tumor–host tissue interactions are inflammatory and immune cell infiltrates. Solid tumors are infiltrated by both innate immunity cells (natural killer cells, macrophages, neutrophils, and phagocytes) and adaptive immunity cells (T lymphocytes, B lymphocytes, and dendritic cells).4Whiteside T.L. The tumor microenvironment and its role in promoting tumor growth.Oncogene. 2008; 27: 5904-5912Crossref PubMed Scopus (1479) Google Scholar Recently, discoveries of the mechanisms by which cancer cells inhibit host antitumor immune response and new immunomodulating therapies have shifted focus toward a search for antitumor immune response components and biomarkers.5Taube J.M. Galon J. Sholl L.M. Rodig S.J. Cottrell T.R. Giraldo N.A. Baras A.S. Patel S.S. Anders R.A. Rimm D.L. Cimino-Mathews A. Implications of the tumor immune microenvironment for staging and therapeutics.Mod Pathol. 2018; 31: 214-234Crossref PubMed Scopus (206) Google Scholar,6Parra E.R. Francisco-Cruz A. Wistuba II, State-of-the-art of profiling immune contexture in the era of multiplexed staining and digital analysis to study paraffin tumor tissues.Cancers (Basel). 2019; 11: E247Crossref PubMed Scopus (73) Google Scholar Tumor-infiltrating lymphocytes (TILs) and their distributions within TME compartments have been reported as potential prognostic and predictive biomarkers in various types of cancer. Studies in clinical and experimental settings have revealed the prognostic role of CD3+, CD4+, CD8+, and FOXP3+ TILs in many types of solid human tumors, such as melanoma, colorectal, breast, lung, bladder, prostate, renal, and hepatocellular carcinomas.5Taube J.M. Galon J. Sholl L.M. Rodig S.J. Cottrell T.R. Giraldo N.A. Baras A.S. Patel S.S. Anders R.A. Rimm D.L. Cimino-Mathews A. Implications of the tumor immune microenvironment for staging and therapeutics.Mod Pathol. 2018; 31: 214-234Crossref PubMed Scopus (206) Google Scholar,7Coussens L.M. Werb Z. Inflammation and cancer.Nature. 2002; 420: 860-867Crossref PubMed Scopus (11186) Google Scholar, 8Fridman W.H. Pages F. Sautes-Fridman C. Galon J. The immune contexture in human tumours: impact on clinical outcome.Nat Rev Cancer. 2012; 12: 298-306Crossref PubMed Scopus (3105) Google Scholar, 9Hussain S.P. Amstad P. Raja K. Ambs S. Nagashima M. Bennett W.P. Shields P.G. Ham A.J. Swenberg J.A. Marrogi A.J. Harris C.C. Increased p53 mutation load in noncancerous colon tissue from ulcerative colitis: a cancer-prone chronic inflammatory disease.Cancer Res. 2000; 60: 3333-3337PubMed Google Scholar, 10Chraa D. Naim A. Olive D. Badou A. T lymphocyte subsets in cancer immunity: friends or foes.J Leukoc Biol. 2019; 105: 243-255Crossref PubMed Scopus (70) Google Scholar Because TILs are represented by several subsets of T and B cells with complex interactions and roles, their assessment in the TME requires taking both functional and spatial aspects into account to understand their roles as major components of antitumor response.11Rosenberg S.A. Restifo N.P. Adoptive cell transfer as personalized immunotherapy for human cancer.Science. 2015; 348: 62-68Crossref PubMed Scopus (1513) Google Scholar A comprehensive study of colorectal cancer (CRC) immunome by Galon et al,12Galon J. Pages F. Marincola F.M. Angell H.K. Thurin M. Lugli A. et al.Cancer classification using the Immunoscore: a worldwide task force.J Transl Med. 2012; 10: 205Crossref PubMed Scopus (610) Google Scholar based on digital image analysis (DIA) of immunohistochemistry (IHC) slides, revealed that the densities of CD3+ and CD8+ TILs in the core tumor and the invasive margin (IM) correlate with the outcome of the disease. This discovery led to a clinically validated Immunoscore indicator, which was superior to the conventional TNM staging system.13Galon J. Mlecnik B. Bindea G. Angell H.K. Berger A. Lagorce C. et al.Towards the introduction of the 'Immunoscore' in the classification of malignant tumours.J Pathol. 2014; 232: 199-209Crossref PubMed Scopus (956) Google Scholar Recent studies have found prognostic and predictive value of high TIL infiltration in triple-negative and human epidermal growth factor receptor 2–positive breast cancer (BC).14Jang N. Kwon H.J. Park M.H. Kang S.H. Bae Y.K. Prognostic value of tumor-infiltrating lymphocyte density assessed using a standardized method based on molecular subtypes and adjuvant chemotherapy in invasive breast cancer.Ann Surg Oncol. 2018; 25: 937-946Crossref PubMed Scopus (32) Google Scholar,15Denkert C. von Minckwitz G. Darb-Esfahani S. Lederer B. Heppner B.I. Weber K.E. Budczies J. Huober J. Klauschen F. Furlanetto J. Schmitt W.D. Blohmer J.U. Karn T. Pfitzner B.M. Kummel S. Engels K. Schneeweiss A. Hartmann A. Noske A. Fasching P.A. Jackisch C. van Mackelenbergh M. Sinn P. Schem C. Hanusch C. Untch M. Loibl S. Tumour-infiltrating lymphocytes and prognosis in different subtypes of breast cancer: a pooled analysis of 3771 patients treated with neoadjuvant therapy.Lancet Oncol. 2018; 19: 40-50Abstract Full Text Full Text PDF PubMed Scopus (916) Google Scholar Naturally, intratumoral T cells are considered an important cornerstone in the emerging concept of Immunogram—a comprehensive assessment of a patient's antitumor response to guide immunotherapies in various cancers.16van Dijk N. Funt S.A. Blank C.U. Powles T. Rosenberg J.E. van der Heijden M.S. The cancer immunogram as a framework for personalized immunotherapy in urothelial cancer.Eur Urol. 2019; 75: 435-444Abstract Full Text Full Text PDF PubMed Scopus (67) Google Scholar, 17Karasaki T. Nagayama K. Kuwano H. Nitadori J.I. Sato M. Anraku M. Hosoi A. Matsushita H. Morishita Y. Kashiwabara K. Takazawa M. Ohara O. Kakimi K. Nakajima J. An immunogram for the cancer-immunity cycle: towards personalized immunotherapy of lung cancer.J Thorac Oncol. 2017; 12: 791-803Abstract Full Text Full Text PDF PubMed Scopus (90) Google Scholar, 18Blank C.U. Haanen J.B. Ribas A. Schumacher T.N. Cancer immunology: the "cancer immunogram".Science. 2016; 352: 658-660Crossref PubMed Scopus (504) Google Scholar The success of these studies has spawned into a search for novel image analytics methods, also benefiting from multiple immunofluorescence techniques, to develop combinatorial image biomarkers for colorectal, prostate, and renal cancer.19Guo C. Zhao H. Wang Y. Bai S. Yang Z. Wei F. Ren X. Prognostic value of the neo-immunoscore in renal cell carcinoma.Front Oncol. 2019; 9: 439Crossref PubMed Scopus (10) Google Scholar, 20Harder N. Athelogou M. Hessel H. Brieu N. Yigitsoy M. Zimmermann J. Baatz M. Buchner A. Stief C.G. Kirchner T. Binnig G. Schmidt G. Huss R. Tissue phenomics for prognostic biomarker discovery in low- and intermediate-risk prostate cancer.Sci Rep. 2018; 8: 4470Crossref PubMed Scopus (18) Google Scholar, 21Nearchou I.P. Lillard K. Gavriel C.G. Ueno H. Harrison D.J. Caie PD: automated analysis of lymphocytic infiltration, tumor budding, and their spatial relationship improves prognostic accuracy in colorectal cancer.Cancer Immunol Res. 2019; 7: 609-620Crossref PubMed Scopus (43) Google Scholar At the core of any method that aims to obtain a better understanding of tumor immune contexture is the task of quantifying the individual immune cell subtypes and their location relative to the tumor cells.8Fridman W.H. Pages F. Sautes-Fridman C. Galon J. The immune contexture in human tumours: impact on clinical outcome.Nat Rev Cancer. 2012; 12: 298-306Crossref PubMed Scopus (3105) Google Scholar,22Hendry S. Salgado R. Gevaert T. Russell P.A. John T. Thapa B. et al.Assessing tumor-infiltrating lymphocytes in solid tumors: a practical review for pathologists and proposal for a standardized method from the international immunooncology biomarkers working group: part 1: assessing the host immune response, TILs in invasive breast carcinoma and ductal carcinoma in situ, metastatic tumor deposits and areas for further research.Adv Anat Pathol. 2017; 24: 235-251Crossref PubMed Scopus (358) Google Scholar,23Hendry S. Salgado R. Gevaert T. Russell P.A. John T. Thapa B. et al.Assessing tumor-infiltrating lymphocytes in solid tumors: a practical review for pathologists and proposal for a standardized method from the international immuno-oncology biomarkers working group: part 2: TILs in melanoma, gastrointestinal tract carcinomas, non-small cell lung carcinoma and mesothelioma, endometrial and ovarian carcinomas, squamous cell carcinoma of the head and neck, genitourinary carcinomas, and primary brain tumors.Adv Anat Pathol. 2017; 24: 311-335Crossref PubMed Scopus (374) Google Scholar The enumeration of immune cells in the TME implies not only accurate detection of the tumor and stroma regions but also a need for a clear and reproducible delineation of the IM.8Fridman W.H. Pages F. Sautes-Fridman C. Galon J. The immune contexture in human tumours: impact on clinical outcome.Nat Rev Cancer. 2012; 12: 298-306Crossref PubMed Scopus (3105) Google Scholar Regardless of whether visual or DIA methods for outlining the IM are applied, the definitions of the IM remain rather ambiguous. An early description of the IM configuration was proposed by Jass et al24Jass J.R. Atkin W.S. Cuzick J. Bussey H.J. Morson B.C. Northover J.M. Todd I.P. The grading of rectal cancer: historical perspectives and a multivariate analysis of 447 cases.Histopathology. 1986; 10: 437-459Crossref PubMed Scopus (498) Google Scholar in 1986, who studied histomorphologic prognostic indicators in rectal carcinoma and defined two different configurations of the IM: expansive (or pushing) and infiltrative. A pushing IM is identified visually when a clear delineation of the tumor and host tissue is possible during examination of the histologic slide. Tumors with an infiltrative IM configuration have a relatively irregular growth pattern in which it is difficult to delineate host tissue from tumor cell aggregates.24Jass J.R. Atkin W.S. Cuzick J. Bussey H.J. Morson B.C. Northover J.M. Todd I.P. The grading of rectal cancer: historical perspectives and a multivariate analysis of 447 cases.Histopathology. 1986; 10: 437-459Crossref PubMed Scopus (498) Google Scholar,25Jass J.R. Love S.B. Northover J.M. A new prognostic classification of rectal cancer.Lancet. 1987; 1: 1303-1306Abstract Full Text PDF PubMed Scopus (536) Google Scholar Many other studies used the IM definition by Mlecnik et al26Mlecnik B. Bindea G. Kirilovsky A. Angell H.K. Obenauf A.C. Tosolini M. Church S.E. Maby P. Vasaturo A. Angelova M. Fredriksen T. Mauger S. Waldner M. Berger A. Speicher M.R. Pages F. Valge-Archer V. Galon J. The tumor microenvironment and Immunoscore are critical determinants of dissemination to distant metastasis.Sci Transl Med. 2016; 8: 327ra26Crossref PubMed Scopus (317) Google Scholar: a 1-mm-wide area around the border separating the host tissue from malignant glands. However, this definition does not provide an explicit explanation of the border; it actually requires an expert's judgment to manually draw it. This remains a source of bias because it leads to interobserver and intraobserver variance in tumors with irregular and highly infiltrative growth patterns, and it surely decreases the capacity of analysis, even if other analysis steps are automated. The only way to achieve a faster manual IM annotation is to simplify the IM shapes, which means leaving out the finer structural details. For instance, in the Immunoscore border, the stromal pathways within the core tumor may not be included as part of the TME. Consequently, the informative power and clinical utility of TILs and other TME-context assays may be underachieved. Recently, Harder et al20Harder N. Athelogou M. Hessel H. Brieu N. Yigitsoy M. Zimmermann J. Baatz M. Buchner A. Stief C.G. Kirchner T. Binnig G. Schmidt G. Huss R. Tissue phenomics for prognostic biomarker discovery in low- and intermediate-risk prostate cancer.Sci Rep. 2018; 8: 4470Crossref PubMed Scopus (18) Google Scholar applied a tissue phenomics approach to search for image-based biomarkers in their study of prostate cancer recurrence prediction. In particular, their DIA step used morphologic operations to automatically delineate tumor gland and stroma areas and subsequently sample the tumor border as a region reaching equally far to both tumor and stroma regions. Another recent study27Schwen L.O. Andersson E. Korski K. Weiss N. Haase S. Gaire F. Hahn H.K. Homeyer A. Grimm O. Data-driven discovery of immune contexture biomarkers.Front Oncol. 2018; 8: 627Crossref PubMed Scopus (17) Google Scholar proposed a data-driven method to discover immune contexture biomarkers; however, it included a manual step of tumor area delineation. In this study, we present a novel set of immunogradient indicators based on a new method of automated grid-based extraction of the tumor-stroma interface zone (IZ). The method first identifies the tumor edge (TE) using a set of explicit rules based on IHC DIA data. Subsequently, the IZ is extracted and ranked by the distance around the TE to allow computing TIL density profiles across the IZ. The indicators, reflecting TIL density gravitation toward the tumor, were independent prognostic factors of better overall survival (OS) in patients with hormone receptor–positive BC and CRC This study was performed in BC and CRC cohorts. The BC cohort consisted of 102 patients diagnosed with early hormone receptor–positive invasive ductal breast carcinoma as described previously.28Laurinavicius A. Green A.R. Laurinaviciene A. Smailyte G. Ostapenko V. Meskauskas R. Ellis I.O. Ki67/SATB1 ratio is an independent prognostic factor of overall survival in patients with early hormone receptor-positive invasive ductal breast carcinoma.Oncotarget. 2015; 6: 41134-41145Crossref PubMed Scopus (14) Google Scholar,29Laurinavicius A. Laurinaviciene A. Ostapenko V. Dasevicius D. Jarmalaite S. Lazutka J. Immunohistochemistry profiles of breast ductal carcinoma: factor analysis of digital image analysis data.Diagn Pathol. 2012; 7: 27Crossref PubMed Scopus (55) Google Scholar Briefly, the patients were women aged 27 to 87 years (median age, 59 years) who underwent surgery in 2007 to 2009 at the National Cancer Institute of Lithuania and were investigated at the Lithuanian National Center of Pathology. The CRC specimens were obtained from 101 patients diagnosed with stage I to III of primary invasive adenocarcinoma and treated by surgical excision without preoperative therapy in 2010 at Vilnius University Hospital Santaros Clinics and investigated at the National Center of Pathology. The CRC cohort included 60 women and 41 men 45 to 89 years old (median age, 70 years). The clinicopathologic and follow-up characteristics of the BC and CRC patient cohorts are summarized in Table 1.Table 1Patient and Tumor Clinicopathologic CharacteristicsCharacteristicBreast cancer cohortColorectal cancer cohortPatients, n102101Age, years Median5970 Range27–8745–89Sex, n (%) Female102 (100)60 (59.0) Male041 (41.0)Follow-up, months Median11266 Range17–1212–75Died, n (%) 5-year follow-up8 (7.8)29 (28.7) 10-year follow-up22 (21.6)Not availableHistological grade, n (%) G123 (22.5)5 (4.9) G248 (47.1)85 (84.2) G331 (30.4)11 (10.9)Tumor invasion stage (pT), n (%) T156 (54.9)5 (4.9) T246 (45.1)19 (18.8) T3062 (61.4) T4015 (14.9)Lymph node metastasis status (pN), n (%) N055 (53.9)57 (56.4) N135 (34.3)24 (23.8) N29 (8.8)19 (18.8) N33 (2.9)1 (1.0) Open table in a new tab The patient studies were approved by and performed in accordance with the guidelines stated by the Lithuanian Bioethics Committee (protocol number 40, April 26, 2007, update September 12, 2017, for the BC patient cohort; protocol numbers L-13-03/1 and L-13-03/2 for CRC patient cohort). Informed written consent was collected from all patients prior to inclusion in the study. Formalin-fixed, paraffin-embedded tissue was used for the study; tumor tissue samples that contained the maximum content of invasive tumor tissue (one formalin-fixed, paraffin-embedded block per patient) were selected. The FFPE tissue sections were cut at 3-μm thickness and mounted on positively charged slides. IHC was performed by Roche Ventana BenchMark ULTRA automated slide stainer (Ventana Medical Systems, Tucson, AZ). Antibodies against cytotoxic T-cell marker CD8 (clone C8/144 B, Dako, Glostrup, Denmark; antibody dilution 1:400) was used followed by use of an ultraView Universal DAB Detection kit (Ventana Medical Systems). The sections were counterstained with Mayer hematoxylin. Positive staining controls were performed in parallel using paraffin-embedded human tonsil tissue. The IHC slides were digitized at ×20 magnification (0.5-μm resolution) using a ScanScope XT Slide Scanner (Leica Aperio Technologies, Vista, CA). DIA of the whole slide images was performed using HALO version 2.2.1870 (Indica Labs, Corrales, NM). Cancer-specific artificial intelligence tissue classifiers were trained using HALO AI module to segment tumor, stroma, and background (consisting of necrosis, artifacts, and glass) and tumor, stroma, lymphoid follicles, necrosis, and background (consisting of glass and artifacts) in BC and CRC, respectively. Subsequently, the Multiplex IHC algorithm version 1.2 in HALO was used to detect and extract coordinates of cytoplasmic CD8+ cells. For quality assurance, all image analysis results were reviewed by a pathologist (A.Laurinavicius): three whole slide images were excluded because of the need to train a specific tissue classifier for the mucinous type of tumor histology. Examples of IHC and DIA analysis output images are presented in Figure 1, A and B . The automated edge extraction uses the pixel-wise classification of the tissue by DIA (Figure 1, A and B). A hexagonal grid (hexagon side, 65 μm) is overlaid in a random position to subsample the DIA results (Figure 1C), as described in previous studies.28Laurinavicius A. Green A.R. Laurinaviciene A. Smailyte G. Ostapenko V. Meskauskas R. Ellis I.O. Ki67/SATB1 ratio is an independent prognostic factor of overall survival in patients with early hormone receptor-positive invasive ductal breast carcinoma.Oncotarget. 2015; 6: 41134-41145Crossref PubMed Scopus (14) Google Scholar,29Laurinavicius A. Laurinaviciene A. Ostapenko V. Dasevicius D. Jarmalaite S. Lazutka J. Immunohistochemistry profiles of breast ductal carcinoma: factor analysis of digital image analysis data.Diagn Pathol. 2012; 7: 27Crossref PubMed Scopus (55) Google Scholar Inside each hexagon, a set V of data variables (IHC-positive and IHC-negative cells counts and tissue class areas) is collected. For robustness to varying hexagon size, the method relies on area fractions inside the grid elements instead of absolute areas. Regardless of how many classes were identified by the classifier, only three area fractions are of interest here: tumor area fraction t, stroma area fraction s, and background area fraction b combined from the remaining classes extracted by DIA. Figure 1D illustrates area fractions (t, s, and b) as red, green, and blue, respectively. For any grid hexagon, h, denote the lookup of data variable v∈Vas v(h), for example, tumor area fraction as t(h) or simply t when h is unambiguous. To identify hexagons with abrupt changes, the change in any v∈Vacross the hexagonal grid is calculated using derivatives along the three hexagonal axes; dxvdenotes derivative of variable v along axis x (at implicit hexagon h); for example, dxt is the change of tumor area fraction along x. Total change is given by the norm of the gradient |∇v(h)|. Thus, the total change in tumor area fraction used for interface extraction is as follows (again omitting h):|∇t|=dxt2+dyt2+dzt2.(1) The numerical implementation of the hexagonal gradient is similar to the hexagonal gradients as calculated according to previous studies.30Coleman S. Scotney B. Gardiner B. Tri-directional gradient operators for hexagonal image processing.J Vis Commun Image Represent. 2016; 38: 614-626Crossref Scopus (10) Google Scholar,31Middleton L. Sivaswamy J. Edge detection in a hexagonal-image processing framework.Image Vis Comput. 2001; 19: 1071-1081Crossref Scopus (64) Google Scholar However, whereas those methods used a linear combination among derivatives to optimize for computation speed, here all derivatives contribute to the gradient to maximize the level of detail extracted. In Figure 1E, |∇t| is overlaid in the red color channel and illustrates how it captures all changes of tumor. For the extraction of the TE, only hexagons where the tumor area changes because of neighboring stroma regions are of interest. Thus, |∇t|is separated into a tumor-stroma part and a tumor-background part by weighting it with the relative changes in stroma and background area fractions. For dxt the relative changes in s and b are as follows:sdxt=dxt|dxs|∑i≠t|dxi|=dxt|dxs||dxs|+|dxb|andbdxt=dxt|dxb|∑i≠t|dxi|=dxt|dxb||dxs|+|dxb|(2) The rationale is that if the amount of background area changes very little across some hexagons, any change in tumor area can be interpreted as being caused by change in stroma area and vice versa. Note that dxts+dxtb=dxt ensures that no information is lost or added but merely separated. The separation weights are similar along y and z, and thus total change |∇t| can be separated by s and b:|∇t|=|∇ts+∇tb|(3) Figure 1F illustrates the separation of tumor changes into tumor-stroma and tumor-background changes on the hexagonal grid using|∇ts|=(dxts)2+(dyts)2+(dzts)2(4) as green and|∇tb|=(dxtb)2+(dytb)2+(dztb)2(5) as blue. To consistently classify the hexagons into tumor, stroma, background, TE, and tumor-background interface, the normalized tumor gradient, denoted ∇t∈[0;1], is used as a probability for whether a hexagon should be part of the TE. The complete set of steps for hexagon classification is as follows: i) Hexagons with abrupt changes in tumor area fraction are identified by testing if ∇t>0.5. For all hexagons for which this is the case, the maximum of ∇ts and ∇tb will subsequently determine whether the hexagon is part of the TE or the tumor-background interface, respectively. ii) Hexagons not deemed part of the TE by step 1 are grouped into tumor, stroma, or background by the maximum of area fractions of t, s, and b because t+ s+b=1, where t,s, and b∈[0;1]. iii) TE invasive regions are identified as hexagons classified as tumor or stroma in step 2 and where the tumor and stroma areas are of similar amount. Specifically, if |t−s|<0.25 the hexagon is deemed part of the invasive areas of TE. Figure 1G shows the result of hexagon classification. Using a simple hexagonal distance transform, each hexagon's shortest distance to the extracted TE was computed. Figure 1H shows the hexagonal distances using a random color for each distance value. To identify the tumor and stroma aspects of the IZ, ranks are established. For any hexagon h,rank(h)def¯¯{ifclass(h)=tumour,rank=distanceifclass(h)=stroma,rank=−distanceifclass(h)=TE,rank=0otherwise,ignorehexagonh(6) The extracted TE has rank 0; inside the tumor, the rank is simply the positive distance from TE, and in stroma, it is the negative distance. The remaining tissue classes (background and tumor-background) are not included in further analyses. The IZ, now consisting of TE with adjacent tumor and stroma tissue, can be defined for different choices of width. Here, IZw denotes IZ consisting of rank interval [−w2;w2], where [ ] is rounding to nearest integer toward 0. Similarly, it is possible to define different widths of central TE by ranks. In this article only, TE = TE1 consisting of r0 and TE3 consisting of ranks [−1; 1] are relevant. Unless otherwise mentioned, TE refers to TE1. Figure 1I shows IZ9; a nine-hexagon-wide IZ with tumor and stroma aspects labeled as red and green hexagons; their color intensity reflects distance to the yellow TE (with an implicit width of 1). Information about cell densities within and across the IZ can be computed by summarizing the hexagonal data values for each rank into rank quantities (Table 2). The rank quantities form a collective interface profile that reveals how cell densities (and other features) vary inside and across the IZ. Examples of CD8+ cell density profiles are given for three tumors in Figure 2. An additional set of cases from both the BC and CRC cohorts illustrating the extracted IZ and immunogradient indicators in various tumor growth patterns and CD8+ density profiles are included in Supplemental Figure S1, Supplemental Figure S10, Supplemental Figure S11, Supplemental Figure S12, Supplemental Figure S2, Supplemental Figure S3, Supplemental Figure S4, Supplemental Figure S5, Supplemental Figure S6, Supplemental Figure S7, Supplemental Figure S8, Supplemental Figure S9.Table 2Summary of Grid Data Variables, Rank Quantities, and Profile IndicatorsValueFormulaStatistical notationGrid data variables Positive cell countsStain No.CD8 AreasArea=No.ofpixelsoftissueclasspixels/μmT, S, and B Area fractionsareaofclasshexagonareat, s, and b Cell densityNo.ofpositivestainsT+SCD8_dRanks and rank quantities riAny rank in IZ of width wri∈IZ9 are [r−4,r−3,…,r3,r4] q(ri)A statistic of hexagons in riSimple q's are mean and sdMean of CD8 density in r2CD8_mean (r2)Mean CD8 in tumor aspectCD8_mean_TProfile indicators CMCM(q)=∑ririq(ri)∑riq(ri)CD8_CM_mean IDID(q)=q(r−1)q(r1)CD8_ID_meanCM, center of mass; ID, immunodrop; IZ, interface zone. Open table in a new tab CM, center of mass; ID, immunodrop; IZ, interface zone. To express the properties of directional variance (gradient) from stroma to tumor aspect, profile indicators can be calculated. Several indicators of cell density variance within and across the IZ were tested for statistical power to predict OS of the patients. Two cell density indicators were found to be the most powerful: Center of mass (CM) was calculated as follows:CM(q)=∑ririq(ri)∑riq(ri),(7) where ri indexes all ranks in the IZ, for example, ri∈[−4;4] for IZ9, and q(ri) denotes choice of
Referência(s)