Context.—

Mitotic count is an important histologic criterion for grading and prognostication in phyllodes tumors (PTs). Counting mitoses is a routine practice for pathologists evaluating neoplasms, but different microscopes, variable field selection, and areas have led to possible misclassification.

Objective.—

To determine whether 10 high-power fields (HPFs) or whole slide mitotic counts correlated better with PT clinicopathologic parameters using digital pathology (DP). We also aimed to find out whether this study might serve as a basis for an artificial intelligence (AI) protocol to count mitosis.

Design.—

Representative slides were chosen from 93 cases of PTs diagnosed between 2014 and 2015. The slides were scanned and viewed with DP. Mitotic counting was conducted on the whole slide image, before choosing 10 HPFs and demarcating the tumor area in DP. Values of mitoses per millimeter squared were used to compare results between 10 HPFs and the whole slide. Correlations with clinicopathologic parameters were conducted.

Results.—

Both whole slide counting of mitoses and 10 HPFs had similar statistically significant correlation coefficients with grade, stromal atypia, and stromal hypercellularity. Neither whole slide mitotic counts nor mitoses per 10 HPFs showed statistically significant correlations with patient age and tumor size.

Conclusions.—

Accurate mitosis counting in breast PTs is important for grading. Exploring machine learning on digital whole slides may influence approaches to training, testing, and validation of a future AI algorithm.

Phyllodes tumors (PTs) of the breast are biphasic neoplasms composed of proliferation of both stromal and epithelial components. According to the latest 5th edition of the World Health Organization Classification of Tumors of the Breast, histologic features to grade PTs include the degree of stromal hypercellularity, stromal mitotic count, stromal atypia and overgrowth, and the nature of the tumor borders.1  Mitotic count is an important histologic criterion for grading and prognostication.2 

Counting mitoses is a routine practice for pathologists evaluating neoplasms microscopically. The current recommendation is to determine the number of mitoses per 10 or 50 high-power fields (HPFs) using the ×40 microscope objective and ×10 eyepiece (overall ×400), with the field diameter and final area of the HPF varying among different microscopes.1  Apart from differences in area of the HPFs, other potential issues include variable field selection for mitotic counting leading to possible misclassification, and the binomial mitotic distribution diminishing its counting accuracy.3,4  To overcome these issues, a larger area can be evaluated.

Digital pathology (DP) is becoming more widely available and has been harnessed to enhance diagnosis and access to subspecialty opinion, promote education, and may also be deployed for remote reporting.5  DP has enabled the development of artificial intelligence (AI) in pathology, through image analysis and machine learning, including working toward open source access.6  AI may be employed as a tool to reduce issues of mitotic counting as well as obviate interobserver reproducibility challenges.7,8  In light of the rapid advancement of AI technology, a suggested framework for AI-mitotic counting could be established.

This study sought to find out if 10 HPFs or whole slide mitotic count would correlate better with clinicopathologic parameters of PTs using DP. We also aimed to determine if this study might aid in providing a baseline for a clinical protocol to count mitoses using AI.

METHODS

Patients and Tumors

Representative slides from 93 consecutive cases of PTs diagnosed between 2014 and 2015 in the Department of Anatomical Pathology, Singapore General Hospital (Singapore), were chosen. Of these, 60 were benign, 31 borderline, and 2 malignant, which were graded based on the World Health Organization guidelines.1  Specifically, stromal hypercellularity and stromal atypia were categorized into mild, moderate, and marked, according to the consensus review by Tan et al.9  A benign PT was defined by mild or moderate stromal hypercellularity, mild or moderate stromal cytologic atypia, occasional mitoses that numbered up to 4 per 10 HPFs, no stromal overgrowth, and pushing margins. A malignant tumor was diagnosed when there was marked stromal cytologic atypia and stromal hypercellularity, stromal overgrowth, high mitotic activity (≥10 per 10 HPFs), and permeative margins; presence of malignant heterologous elements indicated a malignant PT. A borderline tumor possessed some but not all the malignant features. The slides were scanned with the IntelliSite Ultra-Fast Scanner (Philips Digital Pathology Solutions, the Netherlands) and viewed with the Image Management System viewer (Philips).

This study was approved by the SingHealth Centralized Institutional Review Board (CIRB Ref: 2018/2581).

Counting Mitoses With Digital Pathology

Whole slide mitotic counting was conducted on every section and only stromal mitotic figures within the tumor were annotated. Histologically, mitotic figures were spiculated and hyperchromatic. Purely hyperchromatic or pyknotic nuclei were ignored. Tumor characteristics were blinded to prevent bias in counting. After completion of mitotic counting, the areas with the most mitoses in 10 HPFs were annotated. Each HPF, corresponding to a digital ×40 objective, was 0.084 mm2 on the Image Management System viewer. Hence, our digital 10 HPFs spanned an area of 0.84 mm2.

The area of the tumor was demarcated in the Image Management System viewer and recorded. Values of mitoses per millimeter squared were used to standardize the comparison between results from whole slides and those from 10 HPFs.

Statistical Analysis

Using R 3.6.1 developed by the R Core Team, with libraries Hmisc and psychometric, appropriate correlations were conducted between mitotic counts (10 HPFs and whole slide mitoses/mm2) and the clinicopathologic parameters (grade, stromal atypia, stromal hypercellularity, age, and tumor size). For the purpose of the analysis, benign, borderline, and malignant tumors were assigned scores of 1, 2, and 3, respectively. Similarly, stromal hypercellularity and atypia were assigned scores of 1, 2, and 3 for mild, moderate, and marked, respectively. As the mitoses reading was not normally distributed, Spearman's correlation was used to analyze with clinicopathologic parameters.

Figure 1 shows an example of mitoses and Figure 2 shows an illustrative representation of the methodology. The time taken to complete the manual evaluation of one whole slide ranged from 20 to 60 minutes.

Figure 1

Mitotic figures displaying hyperchromasia and spiculation, as indicated in the green circles at ×400 magnification. Routine hematoxylin and eosin–stained sections were used.

Figure 1

Mitotic figures displaying hyperchromasia and spiculation, as indicated in the green circles at ×400 magnification. Routine hematoxylin and eosin–stained sections were used.

Figure 2

An illustrative flow of the methodology. Whole slide mitotic counting was accomplished by manually scanning through the entire slide frame by frame. For mitotic count per 10 high-power fields (HPFs), we chose 10 HPFs with the maximal number of mitoses. For whole slide mitotic count, demarcation was done on the Image Management System viewer to obtain the tumor area. For both whole slide mitotic count and 10 HPFs count, they were converted to mitoses per millimeter squared to standardize the comparison. Routine hematoxylin and eosin–stained sections were used.

Figure 2

An illustrative flow of the methodology. Whole slide mitotic counting was accomplished by manually scanning through the entire slide frame by frame. For mitotic count per 10 high-power fields (HPFs), we chose 10 HPFs with the maximal number of mitoses. For whole slide mitotic count, demarcation was done on the Image Management System viewer to obtain the tumor area. For both whole slide mitotic count and 10 HPFs count, they were converted to mitoses per millimeter squared to standardize the comparison. Routine hematoxylin and eosin–stained sections were used.

RESULTS

All 93 PT cases were included in the statistical analysis. The mean age of the cohort was 42.5 ± 12.7 years (mean ± SD years). Of 93 cases, 60 (64.5%) were benign tumors, followed by 31 (33.3%) borderline tumors and 2 (2.2%) malignant tumors. In terms of atypia, 65 (70.0%) PT cases displayed mild atypia, followed by 27 (29.0%) moderate atypia and 1 (1%) marked atypia. For stromal cellularity, 50 (53.8%) tumors displayed moderate hypercellularity, followed by 31 (33.3%) mild hypercellularity and 12 (12.9%) marked hypercellularity. The mean size of the tumors was 4.6 ± 3.3 cm3 (mean ± SD cm3).

After adjusting for area discrepancies with mitoses per millimeter squared to ensure a fair comparison, the association of mitotic counts with clinicopathologic parameters is presented in the Table. Both whole slide counting of mitoses and 10 HPFs had similar statistically significant correlation coefficients with grade, stromal atypia, and stromal hypercellularity. Neither whole slide mitotic counts nor mitoses per 10 HPFs showed statistically significant correlation with patient age and tumor size.

Correlations of Mitotic Counts Per 10 High-Power Fields (HPFs) and Mitotic Counts on Digital Whole Slides With Clinicopathologic Parameters

Correlations of Mitotic Counts Per 10 High-Power Fields (HPFs) and Mitotic Counts on Digital Whole Slides With Clinicopathologic Parameters
Correlations of Mitotic Counts Per 10 High-Power Fields (HPFs) and Mitotic Counts on Digital Whole Slides With Clinicopathologic Parameters

An inferential analysis was conducted on the correlations of 10 HPFs and whole slide counting. There was a strong correlation between 10 HPFs and whole slide counting (correlation = .794; R2 = 0.63; P < .001; 95% CI, 0.270–0.373).

DISCUSSION

An inferential analysis revealed strong correlation between 10 HPFs and whole section mitotic counting using digital slides, suggesting their interchangeability. This was further supported by their statistically similar significant correlations with grade, stromal atypia, and stromal hypercellularity. Mitotic counts in 10 HPFs represent a key parameter in PT and cancer grade assignments. Bonert and Tate3  commented that pathologists commenced counting when they found a mitosis and counted areas they perceived as the most mitotically active, which could lead to misclassification, depending on tumor type. They showed the current practice of counting mitoses had high misclassification rates of 9% to 16%, which were reduced when a larger area was counted.3  Therefore, it was also possible that 10 HPFs may not optimally reflect the true overall mitotic count that contributes to histologic grading, a view supported by Meyer et al.4  Based on the results by Bonert and Tate,3  15.0 mm2 achieved a 96% correct classification and further increment in area did not improve mitotic count classification.3  Bonert and Tate3  went on to conclude that 10 HPFs did not reflect a good standard sample area due to varying microscope fields. However, our methodology allowed accurate and standardized 10 HPFs to be chosen in a given slide, hence reducing the possibility of misclassification. In our study, the 10 HPFs were selected after the entire slide was scrutinized for mitoses, suggesting this process facilitated accurate designation of the relevant 10 HPFs. Additionally, our results affirm that maximal rather than a random mitotic count per 10 HPFs is the appropriate parameter for grading purposes.

DP has allowed precise annotations, which made counting convenient for mitoses found within a precise and standardized area. In this study, DP enabled us to count the mitoses in a systematic manner frame by frame. When discussing with colleagues, DP ensured the same area could be recalled for combined review or even in separate locations. Therefore, DP has offered many advantages throughout this study, including making mitosis counting subjectively easier.

While the authors did not train, test, or validate any machine learning algorithm, we believe that this study may help to establish a framework for this purpose. With AI, counting mitoses in whole slides should be feasible. After determining mitotic counts in the whole slide, the AI system may also select the area with the highest mitotic count within a standardized area, such as 10 HPFs. Mitotic counts can be provided over a larger and smaller area, with standardization achieved via enumeration as counts per millimeter squared. This can circumvent the issue of sampling and selection challenges in mitotic counting. To the best of our knowledge, no clinically used protocol has been established for mitotic counting with AI technology and most studies have focused on detection of mitotic figures. Our study suggests potential benefit in mitotic counting over a larger area, which may be enabled and facilitated by AI, or in selecting the most appropriate 10 HPFs with the maximal mitotic counts that can contribute to histologic grading of PTs and other tumors. With machine learning, the most mitotically active area can be identified and we could also standardize the area for mitotic counts in the future. Moreover, AI may serve as a springboard for developing a more accurate grading system.

If AI can recognize and count mitoses over the entire slide, the pathologist may merely need to select the slide for mitotic count determination. This can free up time for the pathologist to assess other pathologic parameters, though it is uncertain how much time may be effectively saved. Cost impact is difficult to evaluate per case, but this could be offset by potentially improved diagnostic accuracy. For instance, if digital mitotic counting could promote accurate grading, therapy could be more tailored. Work is currently in progress to develop an AI platform to count mitoses.

Limitations of our study include the evaluation of only PTs, with very few of malignant grades. Also, representative slides of PTs were used, which may not address mitotic count variation across different slides. This could explain the lower than usual mitotic count observed in the malignant samples. There may also be selection bias by annotating only a single set of 10 HPFs area for analysis. In addition, the authors acknowledge the Image Management System viewer's digital HPF is small, hence millimeters squared was used for comparability. The small HPF of digital pathology may also imply the need for a larger area and number of HPF required for digital assessment. Last, time and cost efficiency analysis of an AI system to assess the whole slide, as well as validation by clinical outcomes, will be needed.

CONCLUSIONS

This study showed that an accurate set of 10 HPFs that yielded a maximal mitotic count can be chosen after evaluating the whole slide. DP makes counting mitoses over a larger area subjectively easier, with the possibility of AI being used as facilitator and enabler. This could influence how we approach training, testing, and validation of future AI algorithms for mitotic counting.

References

References
1.
WHO Classification of Tumours Editorial Board
.
Breast Tumours: WHO Classification of Tumours. 5th ed, Volume 2
.
Lyon
:
International Agency for Research on Cancer;
2019
.
2.
Tan
PH,
Thike
AA,
Tan
WJ,
et al.
Predicting clinical behaviour of breast phyllodes tumours: a nomogram based on histological criteria and surgical margins
.
J Clin Pathol
.
2012
;
65
(1)
:
69
76
.
3.
Bonert
M,
Tate
AJ.
Mitotic counts in breast cancer should be standardized with a uniform sample area
.
Biomed Eng Online
.
2017
;
16
(1)
:
28
.
4.
Meyer
JS,
Cosatto
E,
Graf
HP.
Mitotic index of invasive breast carcinoma. Achieving clinically meaningful precision and evaluating tertial cutoffs
.
Arch Pathol Lab Med
.
2009
;
133
(11)
:
1826
1833
.
5.
Montalto
M.
An industry perspective: an update on the adoption of whole slide imaging
.
J Pathol Inform
.
2016
;
7
:
18
.
6.
Janowczyk
A,
Madabhushi
A.
Deep learning for digital pathology image analysis: a comprehensive tutorial with selected use cases
.
J Pathol Inform.
2016
;
7
;
29
.
7.
Veta
M,
van Diest
PJ,
Jiwa
M,
Al-Janabi
S,
Pluim
JPW.
Mitosis counting in breast cancer: object-level interobserver agreement and comparison to an automatic method
.
PLoS One
.
2016
;
11
(8)
:
e0161286
.
8.
Malon
C,
Brachtel
E,
Cosatto
E,
et al.
Mitotic figure recognition: agreement among pathologists and computerized detector
.
Anal Cell Pathol (Amst)
.
2012
;
35
(2)
:
97
100
.
9.
Tan
BY,
Acs
G,
Apple
SK,
et al.
Phyllodes tumours of the breast: a consensus review
.
Histopathology
.
2016
;
68
(1)
:
5
21
.

Author notes

The authors have no relevant financial interest in the products or companies described in this article.