Introduction

Cervical cancer is one of the most common contributor to the global burden of disease. In 2018, approximately 90% of the 570,000 new cases and 311,000 mortalities have been reported to occur in under developed and developing nations, where systematic screening does not exist or has been poorly implemented. [1]

Invasive squamous cell carcinoma of cervix is preceded by the development of dysplastic changes which gradually progress to HSIL and invasive carcinoma over a duration of two decades. Because there are few symptoms during this preinvasive period, there is a wide window of opportunity during which identification and intervention are still possible. The main aetiological cause is unquestionably persistent infection by high-risk human papillomavirus (HPV) genotypes, particularly types 16 and 18, although tobacco use, immunosuppression, and poor genital hygiene are considered important cofactors. [2]

Since its introduction in the 1940s, the Papanicolaou smear has emerged as a highly efficient screening tool to detect precancerous and cancerous lesions based on the nuclear and cytoplasmic characteristics. However, certain limitations exists such as the lack of human resources to screen the slides, time-consuming, and interobserver variability. Moreover, the indeterminate categories like ASC-US (Atypical Squamous Cells of Undetermined Significance) is quite challenging. [3, 4] The drawbacks of this hierarchy have led to a growing trend toward computer-assisted image analysis and machine learning, which promise a diagnostic framework that is both repeatable and observer-independent. [5]

The AI algorithms currently in use, depend on the nuclear characteristics like size, shape, colour or deep learning based image categorization. Lack of uniformity across various imaging platforms, and limited interpretability makes its use challenging. [6-8] Alterations in the nuclear chromatin architecture is one of the most significant changes that occur as the cell transitions through various grades of dysplasia and eventually carcinoma, [9] This objective quantifiable parameter has been less frequently evaluated, more so in cervical carcinogenesis [8]. Gray-Level Co-occurrence Matrix (GLCM) texture analysis is a quantifiable parameter which reflects chromatin organization by measuring the spatial connexions between pixel intensities. This quantifies variations in chromatin heterogeneity, entropy, and local contrast which is not appreciated at routine light microscopy [8]. Researchers have investigated the utility of GLCM in various imaging applications, its role in cervical cytology has not been adequately explored.

Hence the present study was planned with the aim to evaluate the role of nuclear morphometric parameters and GLCM-based texture features in conventional cervical Pap smears. If found useful, these novel parameters would improve discrimination between normal, precancerous and cancerous lesions of cervix and also provide objective adjunct image biomarkers which can be integrated into future computer-assisted and AI-enabled cervical cancer screening algorithms.

Material and Methods

This was a cross-sectional diagnostic analytical study conducted at the department of pathology, at a tertiary care hospital, over a duration of one year from January 2023 to December 2023. The method of sampling was convenient sampling technique.

Figure 1: Image showing the methodological workflow for GLCM texture analysis on cervical cytology smears utilizing the ImageJ software.A: Shows the interface of Image J software. B: Shows the Pap smear image used for analysis.C: Generation of Binary image. D: Setting thresholds and segmentation. E: Morphometric assessment using Analyze particles function. F: Analysing Gray-Level Co-occurrence Matrix (GLCM) Texture using GLCM plugin

The study was conducted as per the ethical principles of declaration of Helsinki and the ethical clearance for the study protocol was obtained from the Institutional ethical committee AIMS,IEC. Written informed consent was obtained from each of the participants at the beginning of sample collection,

The cervical smears from women aged 25 years to 60 years, attending outpatient department of gynaecology were obtained by the clinician following standard sampling protocol, fixed in 95% ethyl alcohol and stained using the conventional Papanicolaou staining technique.

Selection of cytology images

The smears were reviewed separately by two independent cytopathologists with more than 10 years’ experience and who were blinded for the diagnosis. After consensus opinion, further evaluation was performed.

Initially we selected six cases each of negative for Intraepithelial Lesion or Malignancy (NILM), Atypical Squamous Cells of Undetermined Significance (ASCUS), Low-grade Squamous Intraepithelial Lesion (LSIL), Atypical Squamous Cells-high grade lesion cannot be excluded (ASC-H), High-grade Squamous Intraepithelial Lesion (HSIL) and Squamous Cell Carcinoma (SCC) diagnosed on pap smear using Bethesda system for reporting cervical cytology. (Figure 2)

Figure 2: Cervical cytology smears (Papanicolaou, x 400) A: Negative for Intraepithelial Lesion or Malignancy (NILM) showing mature intermediate squamous cells with normochromatic nuclei and regular nuclear membrane. B: Atypical squamous cells of undetermined significance (ASCUS) showing mild nuclear enlargement but regular nuclear membrane and texture. C: Low-grade Squamous Intraepithelial Lesion (LSIL). Mild nuclear enlargement, nuclear membrane irregularity and mildly coarse chromatin D. High-grade Squamous Intraepithelial Lesion (HSIL) showing marked irregularity in nuclear membrane and coarse chromatin along with increased nuclear-to-cytoplasmic ratio.

Slides were visualized and images captured using a trinocular, digital microscope connected to a camera under 40× objective magnification (400× total magnification). Measures were taken to maintain uniformity exposure, white balance settings and illumination during image capture. High resolution, uncompressed images were saved in a dedicated workstation for further image analysis.

Selection of Region of Interest (ROI)

From each case approximately 4 to 5 squamous epithelial cells were selected by the researchers who were blinded for the quantitative image analysis, for nuclear morphometric and GLCM analysis. Cells with well-preserved morphology, without any overlap, and demonstrating crisp nuclear details, and uniform staining characteristics were selected for the analysis. Cells with poor fixation, drying artefact, details obscured by inflammation or degeneration were excluded from the analysis. To minimize sampling bias, multiple nuclei were selected across various microscopic fields and the mean of all the analysed nuclei was taken as the final value.

A total of 166 cells were analysed among the various study groups which included 64 from NILM, 26 diagnosed as ASCUS, 24 diagnosed as LSIL and 52 which were diagnosed as HSIL/SCC.

Pre-processing of images

For morphometric and texture analysis, Image J software was used. Each nuclei represented region of interest (ROI) for both nuclear morphometric and GLCM analysis. Normalization of background and contrast optimization was done to reduce illumination variations and maintain good chromatin architecture respectively for each image. Image scale was calibrated using standards. Each nuclei was subjected to manual segmentation. Further it was converted to 8-bit grayscale image before preforming extraction. The parameters assessed for nuclear morphometric included area, ferret and perimeter. Similarly for GLCM, the parameters assessed were Angular Second Moment (ASM), Inverse Difference Moment (IDM), Contrast, Correlation, and Entropy.

Processing of images for nuclear morphometric analysis

Manual segmentation of nuclear boundaries was done followed by the analysis of following nuclear morphometric parameters: i) nuclear area which represented the total number of pixels inside the nuclear membrane and was expressed in square micrometres (µm²). ii) Nuclear parameter represented the total length of the nuclear membrane expressed in micrometres. iii) Feret diameter represented the longest distance between any two points on the nuclear membrane. These parameters basically reflect the nuclear geometry and size which in turn are the know indicators or dysplastic and malignant changes.

Texture analysis using Grey-Level Co-occurrence Matrix (GLCM)

Gray-Level Co-occurrence Matrix (GLCM) is an objective tool for determining the chromatin organization and nuclear texture.

As stated earlier the segmented nuclei were transformed into 8-bit grayscale format, which generated pixel intensity values ranging from 0 (black) to 255 (white).

In this analysis, a GLCM is produced for each nucleus by computing the rate at which pairs of pixels with specific gray-level values occurred at a distinct spatial relationship within the nuclear image. The GLCM analysis uses a higher order statistical method to quantify spatial relationship among the adjacent pixels which in turn is reflective of the chromatin distribution pattern. The texture Analyser plugin of Image J software was utilized to obtain the texture features.

Following GLCM features were determined: i) Angular Second Moment (ASM) which indicated textural uniformity. ii) Inverse Difference Moment (IDM) indicates local homogeneity. iii) Contrast indicates local intensity variation.iv) Correlation indicates the linear dependence amongst adjacent pixel intensities. v) Entropy represents randomness and disorder in the chromatin.

To ensure measurement reproducibility, 10% of the randomly selected images were subjected to re-analysis after a duration of 2 weeks. The average of the measurements were used to analyse intra-observer reproducibility.

Statistical analysis

Data was entered in Microsoft excel and the statistical analysis was done using SPSS software latest version. Continuous variables were expressed as mean ± standard deviation (SD). Categorical variables were expressed as frequencies and percentages. Comparisons among various groups were performed using one-way Analysis of Variance (ANOVA). The statistical significance level was set at p < 0.05.

Results

The present study included analysis of 166 cells. These cells included 64 from NILM, 26 diagnosed as ASCUS, 24 diagnosed as LSIL and 52 which were diagnosed as HSIL/SCC. (Table 1)

Nuclear morphometry: Comparison of nuclear morphometry parameters across various cytodiagnostic categories is shown in Table 2.

Nuclear area:

In the present study, though variation was observed in the mean nuclear areas across the various diagnostic categories, it was not statistically significant (p=0.160). A marginal increase in mean nuclear area was noted for ASCUS as compared to NILM (2622.50 ± 1340.18 µm² vs 2543.20 ± 1863.24 µm² respectively). Nevertheless, a gradual reduction in the mean nuclear areas was noted for LSIL and HSIL/SCC group (2222.13 ± 413.43 µm² vs 1994.75 ± 1256.16 µm² respectively).

Nuclear Perimeter:

In the present study, there was a statistically significant difference in the perimeter among NILM, ASCUS, LSIL and HSIL/SCC categories (p = 0.032). ASCUS category showed highest mean perimeter followed by NILM (207.24 ± 96.28 µm versus 180.97 ± 57.98 µm). Importantly, mean perimeter showed a progressively decreasing trend with LSIL perimeter of 173.74 ± 15.68 µm, and HSIL/SCC perimeter as 172.29 ± 45.31 µm.

Ferret

In the present study, there was a progressive reduction in Feret diameter from ASCUS to HSIL/SCC; however, this difference was not statistically significant (p = 0.095). Feret diameter for NILM was found to be 55.60 ± 18.03 µm, whereas Feret diameter for ASCUS was slightly higher value (59.08 ± 12.93 µm). LSIL demonstrated lower values of 54.20 ± 4.85 µm similar to HSIL/SCC which showed Feret value of 50.58 ± 14.25 µm).

Overall to summarise, there was a statistically significant difference in nuclear perimeter among NILM, ASCUS, LSIL and HSIL+SCC categories (p 0.032). However other nuclear morphometric parameters like nuclear area and Feret did not show any difference among various groups.

GLCM texture analysis: Comparison of GLCM Textural parameters across various cytodiagnostic categories is shown in Table 3.

Angular Second Moment (ASM)

In the present study, there was a statistically significant difference in Angular Second Moment acrossNILM, ASCUS, LSIL and HSIL/SCC categories (p < 0.001). Highest value of ASM (0.9831 ± 0.0028) was observed for NILM category, which indicates homogeneity in chromatin texture in this category. In contrast, the HSIL/SCC category showed low ASM values (0.5804 ± 0.1605), which indicate chromatin heterogeneity. Indeterminate values obtained for ASCUS and LSIL reflected a gradual progression from normal to disorganised chromatin pattern.

Contrast

In the present study, showed a statistically significant difference across various diagnostic categories from NILM, ASCUS, LSIL and HSIL/SCC categories (p < 0.001). NILM demonstrated lowest contrast (82.70 ± 10.05), as compared to HSIL/SCC which showed highest values (2469.10 ± 1302.66), Contrast represent local variation in chromatic heterogeneity and in the present study the findings obtained reflects the occurrence of marked chromatin irregularity and nuclear atypia across the categories. Intermediate values of contrast for ASCUS and LSIL indicate the progressive increase in chromatin variability with increasing severity of epithelial dysplasia.

Correlation

In the present study, there was a statistically significant difference in correlation across various diagnostic categories from NILM, ASCUS, LSIL and HSIL/SCC categories (p < 0.001). Higher values if correlation were observed for NILM as compared to HSIL/SCC groups (0.00219 ± 0.00084, vs 0.00008 ± 0.00001) respectively. Intermediate values of contrast for ASCUS and LSIL indicate advanced chromatin disruption with increasing severity of epithelial dysplasia.

Inverse Difference Moment (IDM/Angular Difference Moment)

In the present study, there was a statistically significant difference in Inverse Difference Moment (IDM), across various diagnostic categories from NILM, ASCUS, LSIL and HSIL/SCC categories (p < 0.001). High IDM values were observed for NILM and ASCUS nevertheless, a successive reduction was observed across LSIL and HSIL/SCC, and the lowest IDM values were noted in HSIL/SCC (0.95275 ± 0.00792). These findings show that there is increase in chromatin heterogeneity with the increase in lesion severity.

Entropy

In the present study, there was a statistically significant difference in entropy, across various diagnostic categories from NILM, ASCUS, LSIL and HSIL/SCC categories (p < 0.001). Lowest values of entropy were observed for NILM (0.1513 ± 0.0819), as compared to HSIL/SCC which showed the highest values (0.7030 ± 0.2513), reflecting shows chromatin disorganisation.

Trend analysis: Progressive Trend across various across various cytodiagnostic categories is shown in Table 4.

The Jonckheere–Terpstra test was applied to determine the trend across various groups according to the increasing grades of severity of atypia. Initially increasing trend followed by a decreasing trend was noted for nuclear area as the lesions progressed after the ASCUS through LSIL, HSIL and SCC. Other parameters in nuclear morphometry like perimeter and Feret diameter did not show any significant trend. As for GLCM parameters, Angular Second Moment showed a decreasing trend indicating less texture uniformity as the disease progressed. Increase trend in contrast was noted reflecting the increase in the higher chromatin heterogeneity and increasing nuclear atypia. Similarly findings were reported for entropy suggesting increasing nuclear complexity and chromatin disorganization with the increasing grades of severity. Decreasing Inverse Difference Moment (IDM) also reflected similar findings.

Discussion

Cervical cancer screening traditionally relies on cytological evaluation, yet interpretation of Pap smears is often limited by subjectivity and considerable inter observer variability. Nuclear morphometric analysis in cytology introduced an element of objectivity in cytology. Further, with the advent of fractal geometry, the complexity and irregularity of the cellular structures were measured. [8-9] Moving ahead, Computer assisted diagnostic algorithms enhanced consistency and reduced observer fatigue, thereby strengthening large scale screening reliability. [6]

Significant advancements has been made in incorporating artificial intelligence and machine learning in diagnostic cytopathology. Earlier, Prieto Bohórquez et al. and Adam et al., used mathematical characterization of nuclear architecture [10, 11]. Later, Win et al. utilizes fully automated screening systems for cellular classification with high diagnostic accuracy. [7] This advancement echoes the shift towards digital pathology and AI driven diagnostic algorithms in routine practice.

Earlier researchers primarily used nuclear segmentation, in automated cytology, whereas recent studies has shown the importance of cytoplasmic segmentation and the resolution of overlapping cells using marker controlled watershed transformation. Consistent with prior reports, the present research supports the integration of mathematical image analysis and computer assisted diagnostic (CAD) systems as a means to enhance cervical cytology, transforming it into a more precise, reproducible, and objective diagnostic approach. [3, 7]

The present study demonstrates that GLCM texture analysis emerge as objective markers for differentiating cervical lesions by reflecting more chromatin texture alterations accompanying cervical neoplastic transformation, whilst nuclear morphometry parameters showed overall less distinguishing ability.

Nuclear morphometry: In the present study, a significant difference was noted in the nuclear parameter from NILM to HSIL/SCC categories. The nuclear perimeter was highest in ASCUS category, which decreased progressively across LSIL, followed by HSIL/SCC categories. Although nuclear area and Feret showed a decreasing trend, the difference was not statistically significant. These findings are consistent with earlier studies. [10] Among the nuclear morphometric parameters evaluated, nuclear perimeter was the only parameter that demonstrated a statistically significant difference across diagnostic categories, whereas nuclear area and Feret diameter showed decreasing trends without reaching statistical significance. Technical aspects like fixation, smear preparation along with overlapping nuclear features among dysplastic grades may have influenced nuclear area values. Interestingly, ASCUS exhibited the highest perimeter values followed by a progressive decline in LSIL and HSIL/SCC. These findings advocate that despite nuclear size being a reliable marker of carcinogenesis, individual nuclear morphometric parameters do not reliably reflect the multifaceted architectural alterations occurring within the dysplastic nuclei. Earlier studies reveal similar findings, wherein researchers observed that the individual morphometric parameters alone are not reliable, however, the combined use of multiple parameters are better markers for reflecting underlying biological changes.[10]

Angular Second Moment (ASM): Alteration in the chromatin architecture reflected as texture is an important hallmark of neoplastic transformation and progression. In normal epithelial cells, the chromatin is uniformly distributed, and as the lesion progresses from LSIL to HSIL/SCC, the architecture is markedly altered and is reflected by heterogenous chromatin distribution. Angular Second Moment (ASM), is a GLCM parameter which reflects chromatin texture uniformity. [8, 13] In the present study, there was a progressive reduction in ASM values from NILM to HSIL/SCC. Similar findings were reported by Win et al., wherein they reported ASM as the strongest discriminator of malignant cells in cervical cytology using computer assisted cytology systems. [7]

Contrast: Contrast provides information about the local intensity changes with respect to adjacent pixels and is a marker of coarseness of chromatin aggregation and irregularity. With increasing grades of dysplasia the chromatin becomes increasingly coarse and similarly, the contrast values also increase with neoplastic transformation and progression. In the present study, there was thirty-fold increase in contrast values from NILM to HSIL/SCC. Similar findings have been reported by other authors in digital cytology studies evaluating cervical lesions. [8, 13]

Correlation: Correlation shows linear dependency among adjacent pixel intensities and is an indicator of preservation of chromatin pattern. In normal epithelial nuclei, neighbouring pixels have similar intensities and hence the correlation values are high. In dysplastic nuclei, the chromatin is randomly distributed and hence the correlation is decreased. Similar findings were reported by Win et al., who observed decreasing correlation values with increasing severity of dysplasia. [8, 13]

Inverse Difference Moment (IDM): Inverse Difference Moment (IDM), is a marker of nuclear homogeneity. In the sequential transformation of carcinogenesis IDM is reduced progressively, from NILM to HSIL/SCC. IDM values are high in NILM and ASCUS reflecting uniform chromatin organization, whereas significantly lower values in HSIL/SCC support increasing chromatin heterogeneity. IDM measures chromatin smoothness and increase in IDM mirrors increase in nuclear irregularity, signifying malignant transformation. These findings suggest that IDM may complement ASM in objectively determining chromatin disorganization at the cellular level. [8]

Entropy: Entropy is a measure of randomness or disorder in a system. Increase in entropy in the nuclei indicates increase in disorganization of the chromatin which is a hallmark of malignancy. In the present study, there was a progressive increase in entropy from NILM to HSIL/SCC and this was statistically significant. The disorganization in chromatin occurring during malignant transformation and progression results from several factors including chromosomal instability, irregular chromatin condensation and abnormal DNA replication. Several other studies have shown increase in entropy with increasing grades of dysplasia in various organ systems like oral, prostate and breast cancer. [8, 9]

Overall GLCM texture parameters showed a specific trend from normal to increasing grades of epithelial dysplasia and malignancies. From NILM through HSIL/SCC, ASM and IDM values showed a decreasing trend, while contrast and entropy showed a progressively increasing trend. The present study demonstrated intermediate values for ASCUS and LSIL thereby emphasising the biological evolution between benign and high grade lesions. These findings reinforce the utility of GLCM texture parameters for quantitative and objective analysis of these grey zone lesions on cervical cytology. Adopting GLCM in screening protocols can influence triage decisions, reduce false negatives and align with national goals of cancer elimination. [7, 8]

The findings of the present study indicate that GLCM texture analysis is a more objective tool and can be used as an adjunct to routine cytological evaluation for screening cervical lesions. The objective nature of these parameters indicates good reproducibility and reduced interobserver variability. Since the parameters are numerical in nature, they can be superlatively incorporated into the computer assisted diagnostic algorithm and artificial intelligence models. The study provides a cost effective and easily interpretable tool that can be easily integrated into the newer AI based screening and diagnostic models without requirement of high end equipment like whole slide imaging. This application paves way for implementation of automated screening system in resource constrained settings in rural India.

Nevertheless, several limitations exists for application of this system in routine reporting which includes standardization in image acquisition, image quality, precision in segmentation, manual selection of area of interest and annotation of large datasets. Furthermore, large scale data validation is must before it is adopted in routine practice. Similarly, single centre study and modest sample size limits its generalization.

References

1. Sung H, Ferlay J, Siegel RL et al. Global cancer statistics 2020: GLOBOCAN estimates of incidence and mortality worldwide for 36 cancers in 185 countries. CA Cancer J Clin. 2021;71(3):209-49.
2. Schiffman M, Castle PE, Jeronimo J, Rodriguez AC, Wacholder S. Human papillomavirus and cervical cancer. Lancet. 2007;370(9590):890-907.
3. Nayar R, Wilbur DC, editors. The Bethesda System for Reporting Cervical Cytology: Definitions, Criteria, and Explanatory Notes. 3rd ed. Cham: Springer; 2015.
4. Davey DD, Neal MH, Wilbur DC, Colgan TJ, Styer PE, Mody DR. Bethesda 2001 implementation and reporting rates: 2003 practices of participants in the College of American Pathologists Interlaboratory Comparison Program in Cervicovaginal Cytology. Arch Pathol Lab Med. 2004;128:1224-9.
5. Janowczyk A, Madabhushi A. Deep learning for digital pathology image analysis: A comprehensive tutorial with selected use cases. Nat Rev Clin Oncol. 2016;13(3):183-98.
6. Elazab N, Gab Allah W, Elmogy M. Deep learning for cervical cancer diagnosis: advances and challenges. Cancer Control. 2024;31:10732748241234567.
7. Win KP, Kitjaidure Y, Hamamoto K, Aung TM. Computer-assisted screening for cervical cancer using digital image processing of Pap smear images. Appl Sci (Basel). 2020;10(5):1800.
8. Haralick RM, Shanmugam K, Dinstein I. Textural features for image classification. IEEE Trans Syst Man Cybern. 1973;SMC-3(6):610-21.
9. Zink D, Fischer AH, Nickerson JA. Nuclear structure in cancer cells. Nat Rev Cancer. 2004;4(9):677-87.
10. Prieto BE, Rodríguez VJO, Correa HSC, Soracipa MMY. Diagnosis of cervical cells based on fractal and Euclidean geometrical measurements: intrinsic geometric cellular organization. BMC Med Phys. 2014;14:2.
11. Adam RL, Silva RC, Pereira FG, Leite NJ, Lorand-Metze I, Metze K. The fractal dimension of nuclear chromatin as a prognostic factor in acute precursor B lymphoblastic leukemia. Cell Oncol. 2006;28(1-2):55-59. doi:10.1155/2006/409593.
12. Amita K, Kalappa P, Rajini T. The usefulness of gray image fractal analysis – An adjunct tool for the diagnosis of breast lesions at cytology. Journal of Microscopy and Ultrastructure. 2023. 10.4103/jmau.jmau_110_22.
13. Castellano G, Bonilha L, Li LM, Cendes F. Texture analysis of medical images. Clin Radiol. 2004;59(12):1061-9.