A Prospective Study to Assess the Utility of Artificial Intelligence-driven Cervical Screening Using Smart Scope versus Conventional Method

ABSTRACT

Background:

Cervical cancer, the fourth most common cancer among women worldwide, often proves fatal and stems from high-risk human papillomavirus infection. Approximately 90% of cervical cancers can be prevented due to the disease’s slow progression, which allows for a 10-year window for the detection and treatment of precancerous lesions. The World Health Organization called for global action toward the elimination of cervical cancer. One of the main strategies for cervical cancer elimination is to achieve screening coverage of at least 70% of women aged 35–45 years and to ensure that 90% of women diagnosed with precancerous lesions or invasive cervical cancer receive appropriate treatment by 2030. One of the main strategies for cervical cancer elimination is to achieve screening coverage of at least 70% of women aged 35–45 years and to ensure that 90% of women diagnosed with precancerous lesions or invasive cervical cancer receive appropriate treatment by 2030.

Aims and Objectives:

The aims of the study were a comparative analysis of conventional Pap smear cytology and artificial intelligence (AI)-based analysis using Smart Scope for screening of cervical cancer.

Materials and Methods:

A prospective, cross-sectional study was conducted among 128 women at a tertiary care center. Participants underwent cervical screening using Smart Scope imaging, followed by AI-based analysis. The findings were compared with conventional methods using appropriate statistical tests.

Results:

In this prospective cross-sectional pilot study involving 128 samples, AI-assisted screening demonstrated a high specificity (96.0%) and negative predictive value (99.2%) for cervical cancer detection, effectively identifying 89.8% of cases as negative for intraepithelial lesion or malignancy. AI enabled rapid image analysis and gynecologist-dependent diagnostic variability and minimized the subjectivity inherent in human interpretation. AI enabled rapid image analysis, reduced gynecologist-dependent diagnostic variability, and minimized the subjectivity inherent in human interpretation.

Conclusions:

AI-based image analysis shows promise as an adjunct tool in cervical cancer screening but lacks the reliability to function as a standalone modality. Further refinement and integration with the clinical context are warranted to improve its screening effectiveness.

INTRODUCTION

Cervical cancer is the fourth most prevalent cancer among women worldwide and remains a leading cause of cancer-related mortality, primarily attributed to high-risk human papillomavirus (HPV) infection.[1,2] Notably, cervical cancer is one of the few malignancies that can be effectively prevented due to its well-documented natural progression and prolonged premalignant phase. Recognizing this, the World Health Organization launched a global initiative in 2020 with the objective of eliminating cervical cancer by 2030.[3] This strategy emphasizes three key pillars: HPV vaccination, high-performance screening programs, and timely detection and treatment of high-grade cervical intraepithelial neoplasia (CIN 2/3) and invasive cancer.[4]

The conventional Pap smear and colposcopy-guided biopsy remain the gold standard for cervical cancer diagnosis, with disease staging determined through a combination of clinical examination and imaging.[5] However, Pap smear results are not immediate, demonstrate relatively low sensitivity (approximately 50-75%) for detecting cervical precancerous lesions, may be compromised by inadequate sample quality (e.g., due to blood or mucus), and frequently necessitate follow-up procedures such as colposcopy or biopsy for confirmation. Effective implementation of these methods necessitates a well-established healthcare infrastructure and highly trained professionals, including pathologists, cytopathologists, and colposcopy specialists.[6] Furthermore, manual screening is inherently subjective and may result in false-negative diagnoses, leading to delayed detection of precancerous or malignant lesions.[5]

In response to these challenges, numerous high-performance cervical cancer screening programs have been developed in recent years.[3] Emerging advancements in artificial intelligence (AI) technology offer significant potential for automated, objective, and standardized detection of cervical cancer and its precursor lesions.[4] AI-based approaches enhance the accuracy and reproducibility of image interpretation, improve diagnostic consistency, and optimize clinical workflows, ultimately streamlining the screening and diagnostic process.[7] Given the high volume, rapid turnaround requirements, and reliance on microscopic and visual evaluation, automation has the potential to significantly improve the efficiency and reliability of cervical cancer screening.

This study seeks to explore the potential of AI-assisted SmartScope as a feasible and efficient screening modality. Specifically, the objective is to evaluate the concordance between AI-driven image analysis and conventional methods by comparing sensitivity, specificity, predictive values, kappa statistics, and Chi-square metrics. We hypothesize that AI-assisted Smartscope can achieve diagnostic performance comparable to that of traditional cytology and clinicocolposcopy, thereby supporting its integration into routine cervical cancer screening workflows.

MATERIALS AND METHODS

A prospective cross-sectional pilot study was conducted to evaluate the role of AI-based image-stained cytology in comparison with conventional Pap smear results for screening detection of cervical abnormalities. The study includes the following diagnostic approaches: (A) conventional Papanicolaou smear – a widely utilized screening test for identifying abnormal cervical cells. (B) Clinicocolposcopy with Swede scoring – a thorough cervical examination using a magnifying instrument, incorporating a scoring system to assess lesion severity. (C) AI-generated assessment: the application of AI enables the analysis of colposcopy images, providing assessments.

The study was conducted at the Department of Obstetrics and Gynaecology, College of Medicine and JNM Hospital in Nadia district, West Bengal, India, over 12 months (April 2024 to March 2025) following approval (Ref. No. F-24/PR/COMJNMH/IEC/23/2102) from the Institutional Ethics Committee of College of Medicine and JNM Hospital.

A total of 200 women were considered for participation in the study. Of these, 184 patients voluntarily agreed to participate after receiving comprehensive information about the research. Each participant provided written informed consent, confirming their willingness to be included in the study upon fully understanding its objectives and procedures. Among the 184 participants, 128 women underwent Smartscope (from Periwinkle Technologies Pvt. Ltd., Pune, India) and also participated in cytological screening. The Smart Scope® (model CX2.0, developed by Periwinkle Technologies Pvt. Ltd., Pune, India) is a portable opto-digital transvaginal device [Figure 1] that connects to a tablet equipped with preinstalled Net4Medix® Version 53 software (also by Periwinkle Technologies Pvt. Ltd., Pune, India) used for storing patient demographic information and cervical images. Periwinkle Technologies developed and validated the AI model (Smart Scope® CX Version 3) used in this study by training it on over 150,000 images captured with the Smart Scope® device across various regions in India. The machine learning algorithm in the software classifies the images into four distinct categories.

Figure 1.

Study flowchart for participant enrollment, screening, artificial intelligence (AI) interpretation, and cytology correlation. Participants were first screened for eligibility and consent, followed by dual screening with SmartScope and conventional cytology. AI categorization into green, amber, or red guided further cytological assessment. AI: Artificial intelligence

The study primarily focuses on patients presenting with various menstrual and reproductive symptoms that may indicate underlying medical conditions requiring thorough evaluation and diagnosis. These symptoms are classified into the following categories: (i) Menstrual irregularities – including abnormal uterine bleeding, intermenstrual bleeding, oligomenorrhea, menorrhagia, and menometrorrhagia. (ii) Genital inflammatory diseases – characterized by chronic leukorrhea, foul-smelling vaginal discharge, vaginal pruritus, and other inflammatory symptoms. (iii) Postcoital bleeding – unexplained bleeding occurring after sexual intercourse. (iv) Postmenopausal bleeding – vaginal bleeding occurring after menopause, which may indicate significant underlying pathology. Women from two different age groups are being studied, and each has been associated with particular reproductive health issues: reproductive age group: this refers to individuals who are typically between the ages of 18 years to 45–50 years (though this can vary depending on individual circumstances if menopause attended earlier). Postmenopausal age group: Women who have undergone menopause and are usually between the ages of 50 and 70 years are included. Women who had not undergone cytological screening or a conventional Papanicolaou smear, those experiencing vaginal bleeding at the time of examination, individuals with clinically evident invasive cancer, women with a history of hysterectomy, pregnant women, and patients who did not provide informed consent were excluded from the study [Figure 1].

Following the Smartscope examination and scoring, demographic parameters including age, parity, last menstrual period, marital status/number of partners, contraceptive use, age at first intercourse, age at marriage, age at pregnancy, total number of pregnancies, history of previous medical conditions, family history of diseases, and personal habits such as hygiene practices and addictions (e.g., smoking and drug use) of each patient were collected.

For patients with abnormal cytology reports, a cervical biopsy was performed. Histopathological examination remains the gold standard for diagnosing cervical cancer. A directed cervical biopsy was conducted in cases where abnormal findings were observed in the Pap smear and colposcopy assessments, ensuring accurate diagnosis and further management.

Once the participants were enrolled in the study, the next step involved preparing them for the procedures. The lithotomy position was utilized, in which the patient lay on their back with their thighs apart and legs bent at the hips and knees, supported by stirrups for optimal access to the pelvic region. After a thorough explanation of the procedure, each patient was positioned comfortably in the dorsal position, and adequate lighting was adjusted to ensure clear visualization of the cervix. To minimize the risk of infection, the vaginal area was cleansed using antiseptics – specifically, cetrimide and chlorhexidine gluconate.

A Cusco’s vaginal speculum was gently inserted to bring the cervix into view, and patients were encouraged to relax. The appearance of the cervix was noted, and a cervical smear was obtained using Ayre’s spatula, which was rotated 360° in both clockwise and counterclockwise directions while maintaining firm contact with the cervix. The spatula was rotated 3–5 times to ensure an adequate sample. The collected material was then evenly spread onto a glass slide and immediately fixed in a bottle containing absolute alcohol.

The cervical smear was stained using the Papanicolaou (Pap) technique and evaluated based on the Bethesda system.[8] Cytology was classified as positive if any of the following were observed: atypical squamous cells of undetermined significance, low-grade squamous intraepithelial lesion (LSIL), high-grade squamous intraepithelial lesion (HSIL), or cells suspicious for malignancy. Smears were classified as negative if they showed only inflammatory changes.

Following the Pap smear, Smartscope (from Periwinkle Technologies Pvt. Ltd., Pune, India) was performed in all participants. A Smart Scope probe was placed to capture images for diagnostic evaluation. The procedure involved obtaining four distinct images to enable a comprehensive assessment of the cervix: initial image – a baseline image of the cervix was captured without enhancements, providing an overview of its general appearance. Image with green filter – A green filter was applied to enhance the visualization of vascular patterns, highlighting abnormal blood vessels associated with precancerous lesions or abnormal tissue growth. Visual inspection with acetic acid – acetic acid was applied to the cervix before capturing an image. Precancerous or abnormal cells exhibited acetowhitening, appearing as white lesions, aiding in the identification of areas that may require further investigation. Visual inspection with Lugol’s iodine (VILI) – Lugol’s iodine solution was applied to the cervix, where normal epithelial cells absorbed the iodine and stained brown or black, while abnormal or malignant cells appeared yellow or unstained, facilitating lesion detection. The severity of cervical lesions following Smartscope examination, using the Swede scoring system, was evaluated on five key parameters, assigning scores between 0, 1, or 2 based on the severity of the findings: acetowhiteness, margin and surface contour, vascular patterns, lesion size, and iodine staining (VILI reaction). Any abnormal findings were systematically documented, and both Papanicolaou smear and clinic-colposcopy examination with Smartscope and AI-generated image analysis were performed for comparison.[9]

Artificial intelligence methodology (SS-artificial intelligence assessment)

Following image capture with the Smart Scope® device, internet connectivity was established to enable analysis via a cloud-based machine learning model. The AI algorithm automatically classified cervical images into one of four predefined diagnostic categories, with corresponding color codes displayed upon selecting the “Assess” button on the tablet interface. These categories included: (i) gsreen: a normal and healthy cervix; (ii) amber: benign conditions including ectropion, cervical polyps, cervicitis, inflammation, infections, or squamous metaplasia, high risk amber (HRA): probable low grade cervical intraepithelial neoplasia (CIN), red: probable high grade CIN or invasive cervical carcinoma. High risk categories such as red and HRA were classified as screen positive by the AI system. All smartscope (SS) AI screen positive flagged cases went for further clinical action.

Statistical analysis

All results were subjected to statistical analysis. MedCalc version 15.8 [Mariakerke, Belgium: MedCalc Software bvba, 2015] was used for descriptive statistics. Descriptive and inferential analyses were conducted using age and Swede score as continuous variables. The correlation between the two methodologies, conventional Pap smear and AI-interpreted results. The kappa coefficient was calculated, and the positive paired samples were analyzed using the Chi-square test.

RESULTS

A total of 128 women were recruited into the study. Table 1 displays descriptive statistics for both age and Swede scores. With a standard deviation of 10.85, the participants’ mean age was 37.57 years (95% confidence interval [CI]: 35.67–39.47). With a median age of 35.0 and an interquartile range (IQR) of 29.5–44.5 years, the age distribution was 19–71 years old. A statistically significant deviation from normalcy was found (P < 0.001).

Table 1.

Descriptive statistics of numerical variables

Parameters	n (sample size)	Minimum	Maximum	Mean	95% CI	Median	SD	25–75 percentile	Normal distribution
Age	128	19.0	71.0	37.57	35.67–39.47	35.0	10.85	29.5–44.5	<0.001
Swede score	128	0.0	8.0	1.20	35.67–39.47	1.0	1.64	0.0–2.0	<0.001

CI: Confidence interval, SD: Standard deviation

With a median of 1.0 and an IQR of 0.0-2.0, the mean Swede score was 1.20 (95% CI: 0.92–1.49). The standard deviation was 1.64, and the score varied from 0 to 8. As would be predicted in a screening sample with low disease frequency, the distribution of Swede scores was deviated significantly from normality (P < 0.001) [Table 1].

Of the 128 cases, 3 (2.3%) had a cytological report of HSIL, 10 (7.8%) of benign. The other 115 (89.8%) cases had negative intraepithelial lesion or malignancy [Table 2]. The majority of samples (78.9%) were classified as green by AI, indicating that AI and traditional Pap smears agree well and AI is generally confident in normalcy. There were no false negatives for HSIL, and all three HSIL instances were marked as either amber (1) or red (2). However, a problem with amber or red false positives is that the AI overcalls a lot of negatives for intraepithelial lesion or malignancy (NILM) and innocuous situations [Figure 2]. The calculated kappa coefficient was 0.184 (95% CI: 0.011–0.357). The CI (0.011–0.357) shows the precision of this estimate. Since the lower bound is just above 0, this suggests some real agreement, but it is still weak.

Table 2.

Distribution of cytology results according to artificial intelligence-predicted Pap smear categories

Pap smear
AI	NILM	Benign	HSIL	Total, n (%)
Green*	95	6	0	101 (78.9)
Amber**	17	2	1	20 (15.6)
Red†	3	2	2	7 (5.5)
Total, n (%)	115 (89.8)	10 (7.8)	3 (2.3)	128

*Healthy cervix, **Benign conditions (e.g., ectropion or polyp), ^†Suspect HGSIL observed. HSIL: High-grade squamous intraepithelial lesion, Benign: Cervicitis, inflammation, polyp, ectropion, squamous metaplasia, infection, atypical squamous cells of indeterminate significance, NILM: Negative for intraepithelial lesion or malignancy, AI: artificial intelligence

Figure 2.

Cervical image assessment across artificial intelligence (AI)-based risk categories and sequential staining. Images captured using SmartScope showing the cervix at baseline (original), after application of acetic acid (post acetic acid), and following Lugol’s iodine staining (post Lugol’s iodine). The images are categorized based on AI risk stratification into green (low risk), amber (intermediate risk), and red (high risk). Characteristic visual changes corresponding to each risk category are evident, supporting AI interpretation

The sensitivity, specificity, positive predictive value, and negative predictive value of AI performance for the detection of HSIL are 66.7%, 96.0%, 28.6%, and 99.2%, respectively [Table 3].

Table 3.

Performance of artificial intelligence for detecting high-grade squamous intraepithelial lesion

Performance Parameter	Value (%)
Sensitivity	66.7
Specificity	96.0
PPV	28.6
NPV	99.2
Total Samples	128

PPV: Positive predictive value, NPV: Negative predictive value

The AI system identified 121 (94.5%) cases as nonred and 7 (5.5%) as red (high-risk) in a dataset of 128 cervical cytology samples. Two (66.7%) of the three HSIL cases found by conventional cytology were marked as red by the AI, while one was overlooked (false negative). A moderate false-positive rate was also caused by the AI’s incorrect classification of five non-HSIL cases as red (2 benign, 3 NILM). The P value found was < 0.001. The results of the Chi-square test showed a statistically significant correlation between cytological diagnoses and AI classifications [Table 4].

Table 4.

Diagnostic distribution of cytology results by artificial intelligence triage category (red vs. nonred)

Pap smear
AI	Benign	HSIL	NILM	Total, n (%)
Nonred*	8	1	112	121 (94.5)
Red**	2	2	3	7 (5.5)
Total, n (%)	10 (7.8)	3 (2.3)	115 (89.8)	128

*Healthy cervix, cervicitis, inflammation, polyp, ectropion, squamous metaplasia, infection, atypical squamous cells of indeterminate significance, **Suspect HGSIL observed. HSIL: High-grade squamous intraepithelial lesion, Benign: Cervicitis, inflammation, polyp, ectropion, squamous metaplasia, infection, atypical squamous cells of indeterminate significance; NILM: Negative for intraepithelial lesion or malignancy, AI: Artificial intelligence

DISCUSSION

AI has seen substantial growth in recent years, particularly in the medical field, where its image-recognition capabilities are being increasingly harnessed for diagnostic applications.[10,11] The healthcare sector, currently grappling with a shortage of trained professionals, regional disparities in care, and increasing workloads, may benefit significantly from AI-driven support systems. Several large-scale studies have been conducted on the use of AI in cervical diagnostics. While much attention has focused on enhancing colposcopy, there is also significant interest in leveraging AI to support cytological analysis and evaluate cervical epithelial changes.[12]

Cervical cancer screening traditionally relies on a tiered protocol: initial cytological or HPV testing, followed by colposcopy and biopsy in cases of abnormal findings. This diagnostic cascade, though effective, is prone to variability and resource constraints, particularly in low-resource settings. AI-based platforms trained to differentiate between LSIL and HSIL offer potential for real-time, standardized assessments that may improve triage efficiency.

The current study highlights both the promise and pitfalls of AI-assisted image interpretation. Notably, several cases reported as NILM by conventional cytology were flagged as amber or red by AI. These false-positive results, while promoting vigilance, carry the risk of unnecessary patient anxiety, additional diagnostic procedures, and potential overtreatment. These misclassification patterns underscore the dual-edged nature of AI integration into clinical workflows and emphasize the need for calibrated thresholds and continuous performance monitoring.

Several limitations must be acknowledged. Smartscope devices used in this research capture only ectocervical images, thereby excluding potential endocervical lesions, which require further evaluation using endocervical brush cytology or curettage. Additionally, AI performance may have been influenced by inconsistencies in image acquisition, including suboptimal focus and varying distances between the camera and the cervix.

Colposcopy-guided biopsy remains the gold standard for diagnosis, but AI-based Smartscope cervical screening holds promise in supplementing cervical cancer screening, particularly in community-based screening in a large population even in the absence of specialist doctor. However, this study highlights the need for well-designed, larger multicentric studies to evaluate diagnostic accuracy, assess clinical feasibility, and refine AI model thresholds.

CONCLUSIONS

A substantially larger and more diverse study population is necessary to draw definitive conclusions about the efficacy and reliability of AI-based screening. At present, its utility in reinforcing low-risk assessments could offer value in resource-limited settings. To strengthen the evidence and support clinical adoption, future research should prioritize well-designed prospective studies, including randomized controlled trials, that directly compare human-performed colposcopy with AI-assisted colposcopy. Emphasis should be placed on evaluating diagnostic accuracy, scalability, cost-effectiveness, and usability, particularly in low-resource and high-burden settings. Such evidence will be crucial for guiding policy decisions and integrating AI tools into cervical cancer screening programs.

Conflicts of interest

There are no conflicts of interest.

Funding Statement

This study was financially supported by the Science and Engineering Research Board, Government of India (SUR/2022/000616).