Development and Clinical Validation of Visual Inspection With Acetic Acid Application-Artificial Intelligence Tool Using Cervical Images in Screen-and-Treat Visual Screening for Cervical Cancer in South India: A Pilot Study Authors: Usha Rani Poli, Anirudh G Gudlavalleti, Jaya Bharadwaj Y, Hira B Pant, Varun Agiwal, GVS Murthy Published: 2024 ABSTRACT AbstractPURPOSEThe burden of cervical cancer in India is enormous, with more than 60,000 deaths being reported in 2020. The key intervention in the WHO's global strategy for the elimination of cervical cancer is to aim for the treatment and care of 90% of women diagnosed with cervical lesions. The current screen-and-treat approach as an option for resource-limited health care systems where screening of the cervix with visual inspection with acetic acid application (VIA) is followed by immediate ablative treatment by nurses in the case of a positive test. This approach often results in overtreatment, owing to the subjective nature of the test. Unnecessary treatments can be diminished with the use of emerging computer-assisted visual evaluation technology, using artificial intelligence (AI) tool to triage VIA-positive women. The aim of this study was (1) to develop a VIA-AI tool using cervical images to identify and categorize the VIA-screen–positive areas for eligibility and suitability for ablative treatment, and (2) to understand the efficacy of the VIA-AI tool in guiding the nurses to decide on treatment eligibility in the screen-and-treat cervical screening program.METHODSThis was an exploratory, interventional study. The VIA-AI tool was developed using deep-learning AI from the image bank collected in our previously conducted screening programs. This VIA-AI tool was then pilot-tested in an ongoing nurse-led VIA screening program.RESULTSA comparative assessment of the cervical features performed in all women using the VIA-AI tool showed clinical accuracy of 76%. The perceived challenge rate for false positives was 20%.CONCLUSIONThis novel cervical image–based VIA-AI algorithm showed promising results in real-life settings, and could help minimize overtreatment in single-visit VIA screening and treatment programs in resource-constrained situations. INTRODUCTION Cervical cancer is the fourth most common cancer among women worldwide, with 570,000 new cases and 311,000 deaths by 2020.1 Nearly 90% of these deaths occur in low-income and middle-income countries. Disparities in the distribution of incidence and mortality rates are due to inadequate access to health services. The cervical cancer burden in India is enormous, with more than 60,000 deaths in 2020, and is a major cause of the lack of organized cervical cancer screening. The key intervention in the WHO's global strategy for the elimination of cervical cancer is to aim for the treatment and care of 90% of women diagnosed with cervical lesions. The recommended screen-and-treat approach option is for a resource-limited health care system where screening of the cervix by visual inspection with acetic acid application (VIA) is followed by immediate ablative treatment by nurses.2 A health worker performing VIA rates the cervical appearance as normal or abnormal, paying particular attention to possible invasive cancers. If the appearance is abnormal, and after excluding invasive cancer, based on the size and position the visible precancerous lesion, it permits the cervical epithelium to be ablated by freezing or heating. Excisions are occasionally required for the treatment of severe lesions. CONTEXT Key Objective Key Objective To develop a novel visual inspection with acetic acid application-artificial intelligence (VIA-AI) tool on the basis of deep learning of AI to apply to triage VIA test-positive women to assist nurses in assessing treatment eligibility in the screen-and-treat approach. To develop a novel visual inspection with acetic acid application-artificial intelligence (VIA-AI) tool on the basis of deep learning of AI to apply to triage VIA test-positive women to assist nurses in assessing treatment eligibility in the screen-and-treat approach. Knowledge Generated Knowledge Generated In this exploratory intervention study, the VIA-AI tool was developed from cervical image bank of local sources. When piloted in a real-life program, this tool showed promising data with clinical accuracy of 76% compared with the gold standard test. In this exploratory intervention study, the VIA-AI tool was developed from cervical image bank of local sources. When piloted in a real-life program, this tool showed promising data with clinical accuracy of 76% compared with the gold standard test. Relevance Relevance Screening of the cervix with VIA is followed by immediate ablative treatment by nurses in the case of a positive test. This approach is adopted in low- and middle-income countries and often results in overtreatment, owing to the subjective nature of the test. Such a novel technology can be successfully integrated as a key adjunct to a screening program, especially in low-resource settings, while maintaining high quality standards. Screening of the cervix with VIA is followed by immediate ablative treatment by nurses in the case of a positive test. This approach is adopted in low- and middle-income countries and often results in overtreatment, owing to the subjective nature of the test. Such a novel technology can be successfully integrated as a key adjunct to a screening program, especially in low-resource settings, while maintaining high quality standards. Preliminary data from our earlier study on VIA-based screen-and-treat cervical screening by trained nurses as a community outreach program in a rural area of Telangana state (India) showed better compliance with cryotherapy for VIA-positive women in a single-visit approach. However, a clear-cut overtreatment rate of seven times was observed.3 Recent (unpublished) data from our ongoing VIA screening and thermal ablation as a service program by nurses also showed good acceptance, but continued to have an overtreatment rate of approximately five times. In a randomized trial in Zambia, an low- and middle-income country, the estimated proportion of overtreatment in a VIA screen-and-treat program showed that nearly half of the women treated had no cervical intraepithelial neoplasia (CIN) and only one third of the treated women had high-grade CIN.4 Hence, the major concern of VIA screen-and-treat program is a low positive predictive value, resulting in overtreatment. The limited availability of expert colposcopists in low-resource settings represents another bottleneck.5 The government of India has currently implemented VIA–based screening method and promoted it under a national program to cover a large population by trained nurses to conduct screening of women in primary health centers.6 However, VIA is highly subjective and has limited accuracy, resulting in over-referral and overtreatment.7 This finding has major implications for the availability of financial, physical, and human resources. Scaling up screening programs for detecting and treating cervical precancers is limited because of the need for intensive quality assurance efforts and difficulty in ensuring adequate provider training and retraining. Therefore, new tools for triaging and guiding the detection of eligibility for ablative treatment are required in single-visit programs. Artificial intelligence (AI) is a computer-performed task that is commonly associated with human intelligence. Deep learning (DL) uses artificial neural networks and their models, which allow machines to imitate how the human brain processes information to learn from large data sets. It was shown that a well-designed and trained DL algorithm could perform classification tasks and make predictions with high accuracy similar to human expert–level performance, thus leading to better health outcomes.8 Asiedu et al9 used a DL AI model in a pocket colposcope device to distinguish CIN from benign tissue and achieved a sensitivity, specificity, and accuracy of 81.3%, 78.6%, and 80.0%, respectively.9 Hu et al10 applied the DL approach to evaluate digitized images of the cervix in an automated process to predict the probability of the image being represented as a precancer or cancer case. The sensitivity of this approach was 91% and established a higher accuracy compared with pap smears in the cohort with a 7-year follow-up.10 Digital VIA performed by a smartphone has a higher discriminative power than naked-eye examination in detecting precancerous lesions, making it an additive value to traditional VIA for improving the diagnosis of cervical precancerous lesions.11 Therefore, this study aimed to develop a novel AI-based tool that can be used to overcome the above-mentioned limitations in triaging VIA test-positive women, assisting nurses in assessing their eligibility for ablative treatment. This would also help to minimize overtreatment at the primary care level. The objectives of this research are: To develop a novel VIA-AI tool on the basis of emerging computer-assisted visual evaluation technology, using DL of AI that can be applied to triage VIA test-positive women in assessing their treatment eligibility. To develop a novel VIA-AI tool on the basis of emerging computer-assisted visual evaluation technology, using DL of AI that can be applied to triage VIA test-positive women in assessing their treatment eligibility. To assess and understand the clinical efficacy of the VIA-AI tool in guiding nurses in deciding treatment eligibility in cervical screening programs. To assess and understand the clinical efficacy of the VIA-AI tool in guiding nurses in deciding treatment eligibility in cervical screening programs. METHODS This study was designed as an exploratory intervention study, carried out in two phases: the development of the VIA-AI algorithm as an application tool and the evaluation of the clinical efficacy of this tool. Ethical approval was obtained from the Institutional Ethics Committee of the Indian Institute of Public Health, Hyderabad, India. Phase I: Development of the AI Algorithm VIA-AI Software The process for development of the VIA-AI consisted of colposcopy, which uses a low-power microscope to visualize cervical changes that are highlighted with the application of exogenous contrast agents such as 3%-5% acetic acid. The system was developed keeping this principle at the forefront and using cervical images of women from local settings to best represent the morphologic changes in the women of this region. The algorithm was developed so that it would identify and categorize the VIA-positive lesions into (1) normal, those that do not require any treatment, or (2) abnormal, that is, (a) those that could go for immediate ablative treatment or (b) to be referred for excisional methods and (c) those with suspected cancers for definitive treatment procedures. Image collection and selection was done as shown in Figure 1. The preprocessing of the speculum reflections and detection of region of interest was undertaken (Appendix 1). Image characterization was done to label them on the basis of the intensity and the margins of acetowhite (AW) areas for the VIA-positive images (Fig 2, Appendix Table A1). To increase the training data set size and generalize the images, multiple image augmentation policies were applied to overcome the data imbalance limitations. Training data sets and testing/validation data sets were randomly selected at the ratio of 80:20 from the total available images. Transfer learning of convoluted neural network models with Efficient net B3 and Efficient V2 models that were selected on the basis of performance on the internal validation data set (Appendix Table A2). The training of training data sets was done by the transfer learning method, where only the final layers of the data sets were trained by keeping the weights of all the remaining layers frozen. Prototype development and cross-validation of the VIA-AI tool was performed, and a real-time automated application was developed for use on mobile phones or tablet devices. After the models were developed, this tool was tested on another separate testing data set consisting of a new random set of 100 images for cross-validation. The main outcome measures and statistical analysis showed an AI prediction accuracy of 78% in identifying correct intensity scores of AW areas, 72% in identifying correct margin scores of AW areas, and 82% in identifying VIA-positives suspicious of invasive cancer. Image collection and characteristics. VIA, visual inspection with acetic acid application. (A) Images with normal features. (B) Images with abnormal features. (C) Images with invasive cancer features. Phase II: To Understand the Clinical Efficacy and Validation of the VIA-AI Tool—A Pilot Study in a Community-Based VIA Screening Program Real-World Data After the AI tool (VIA-AI) was developed, it was piloted at an ongoing opportunistic screen-and-treat visual cervical screening program by trained nurses with 3-12 years of experience in women older than 25 years in community-level cancer screening clinics of Telangana, India. The recruitment of women for field-testing using the VIA-AI tool began between December 2021 and May 2022. VIA screening was performed, followed by colposcopy examination with the Swedes score assessment. Later, all women underwent colposcopy examination with a magnification of 6× of the video colposcope, and images from the video monitor were captured using Android smartphone cameras with 20 megapixel resolution held parallel at a distance of two feet. Each woman had one image captured 1 minute after acetic acid application, which was labeled with their identification number. The data and images were transferred into the database, and cervical biopsy for histologic confirmation was performed on women with a Swede score greater than four. The captured images were run in the VIA-AI software-blinded manner with clinical details, and AI prediction results were reported. Data analysis was performed to estimate the accuracy and efficacy of VIA-AI prediction. Statistical Analysis The primary outcome measure was to estimate the accuracy of AI prediction (using the developed AI tool named VIA-AI) against that of the colposcopy Swede score >4 as a reference standard, and histologic detection rates of cervical lesions CIN1+ as the gold standard. These measures were used, as they are considered the end points for advising treatment for VIA-screen–positive women in a screen-and-treat method for cervical cancer screening programs in low-resource settings. This study compared the diagnostic accuracy of the developed VIA-AI tool to detect CIN 1+ (precancerous lesions and cancer) among VIA-positive women using the developed VIA-AI tool. The sensitivity, specificity, and positive and negative predictive values were calculated using 95% CIs. A receiver operating characteristic (ROC) curve (sensitivity against 1 – specificity) was plotted to evaluate the diagnostic ability of the AI model. Calculations were performed using the statistical analysis software package (Stata 14). True-positive images showed acetowhitening in the VIA image, whereas true-negative images showed little to no visible whitening in the VIA images. False-positive images classified as negative by histopathology and colposcopy but positive by the algorithm were due to image noise. False-negative results showed no acetowhitening on the VIA image with lesion. Errors in incorrect predictions were also assessed. RESULTS Approximately 3,010 eligible women were sequentially recruited and screened to undergo the VIA test, colposcopy, image capture, and guided cervical biopsy for a Swede score of >4. The data of 2,052 women consisting of cervical images were included in the data analysis, and those with missing cervical images were excluded. The mean age of the screened women was 39.53 ± 9.43 years, the minimum age was 25 years, and the maximum age was 70 years (Table 1). TABLE 1Distribution of AgeAge, YearsNo. (%)25-29355 (17.3)30-34294 (14.3)35-39390 (19)40-44299 (14.6)45-49349 (17)50-59293 (14.3)>6072 (3.5) Distribution of Age Of the analyzed data, the proportions of cervical lesion detection on the basis of the VIA test, colposcopy Swede score, final diagnosis of CIN1+, and VIA-AI tool prediction performance were 267 (13.18%), 183 (9.04%), 88 (4.29%), and 103 (5.02%), respectively (Table 2). Thus, AI reduced the false-positive rate of VIA-positive women by 61%. Among the 267 VIA-screen–positive women, 103 (38.58%) cervical lesions were correctly diagnosed by the AI tool, 85 (46.45%) of 183 colposcopy Swede scores >4, and 55 (62.50%) of 88 in the final diagnosis of CIN1+ were correctly predicted by the tool. The overall sensitivity values of AI through the three gold standards VIA test, Swede score, and final diagnosis were 38.58% (95% CI, 32.71 to 44.70), 46.45% (95% CI, 39.06 to 53.95), and 62.50% (95% CI, 51.53 to 72.60), respectively, while the specificity values were 100% (95% CI, 99.79 to 100.00), 99.04% (95% CI, 98.48 to 99.43), and 97.56% (95% CI, 96.77 to 98.19), respectively (Table 3). The specificity values indicated as normal cases were found to be high, at approximately 97%-100%. Therefore, AI tools can be considered as systems that can identify and predict cervical lesions with significant improvement in positive predictive value compared with other standard clinical methods. TABLE 2Distribution of Participant's Results–Based AI and Different Gold StandardAI PredictionGold Standard/Reference Test, No. (%)Final DiagnosisColposcopy Swede ScoreVIA TestPositive (CIN1+)Negative (normal)Positive (score >4)Negative (score <4)PositiveNegativePositive55 (62.50)48 (2.44)85 (46.45)18 (0.96)103 (38.58)0Negative33 (37.50)1,916 (97.56)98 (53.55)1,851 (99.04)164 (61.42)1,785 (100)Total881,9641831,8692671,785Abbreviations: AI, artificial intelligence; CIN, cervical intraepithelial neoplasia; VIA, visual inspection with acetic acid application. Distribution of Participant's Results–Based AI and Different Gold Standard Abbreviations: AI, artificial intelligence; CIN, cervical intraepithelial neoplasia; VIA, visual inspection with acetic acid application. TABLE 3Sensitivity, Specificity, and ROC Area Under Curve of the AI With Different Gold StandardsAI PredictionGold Standard/Reference TestFinal DiagnosisColposcopy Swede ScoreVIA TestSensitivity (95% CI)62.50 (51.53 to 72.60)46.45 (39.06 to 53.95)38.58 (32.71 to 44.70)Specificity (95% CI)97.56 (96.77 to 98.19)99.04 (98.48 to 99.43)100 (99.79 to 100.00)PPV (95% CI)53.40 (43.30 to 63.29)82.52 (73.79 to 89.30)100 (96.48 to 100.00)NPV (95% CI)98.31 (97.63 to 98.83)94.97 (93.91 to 95.90)91.59 (90.26 to 92.78)ROC area0.75850.88750.9579Abbreviations: AI, artificial intelligence; NPV, negative predictive value; PPV, positive predictive value; ROC, receiver operating characteristic; VIA, visual inspection with acetic acid application. Sensitivity, Specificity, and ROC Area Under Curve of the AI With Different Gold Standards Abbreviations: AI, artificial intelligence; NPV, negative predictive value; PPV, positive predictive value; ROC, receiver operating characteristic; VIA, visual inspection with acetic acid application. Among the VIA-screen–positive women, the sensitivity value of AI compared with the final diagnosis and Swede score was 64.71 (53.59-74.77) and 47.49 (39.99-55.07; Table 3), respectively. This shows that AI correctly identified true-positive cases compared with the final diagnosis and Swede scores by 64.7% and 47.5%, respectively. The proposed AI tool has a more than 60% chance of predicting the correct diagnosis among women who test positive for VIA screening. Of the 2,052 interpretations, 402 (20%) were incorrect. The most common reasons for incorrect interpretation were thick mucus (24%), inflammation, squamous metaplasia (18%), and glare spots (15.5%; Table 4; Fig 3). Of the 265 VIA-positive cases, the major reasons for incorrect AI prediction were ectropion (24.4%), inflammation and squamous metaplasia (18.1%), and poor image quality (16.5%). The current challenge is to reduce the number of false positives in the groups that require ablative treatment. TABLE 4Reasons for VIA-AI Misinterpretation in Overall and AI-PositivesReasons for Wrong InterpretationOverall (n = 2,052), No. (%)VIA-AI–Positives (n = 265), No. (%)Mucus99 (24)22 (18.1)Blood32 (7.9)9 (7.4)Ectropion43 (10.6)30 (24.8)Inflammation and squamous metaplasia82 (20)22 (18.1)Atrophic changes10 (2.5)2 (1.66)CIN17 (4.2)0Glare spots63 (15.5)16 (13.2)Poor quality of image61 (15)20 (16.5)Total407 (20)121 (45.6)Abbreviations: AI, artificial intelligence; CIN, cervical intraepithelial neoplasia; VIA, visual inspection with acetic acid application. Reasons for VIA-AI Misinterpretation in Overall and AI-Positives Abbreviations: AI, artificial intelligence; CIN, cervical intraepithelial neoplasia; VIA, visual inspection with acetic acid application. Examples of wrong interpretations as AI-positives. AI, artificial intelligence. DISCUSSION AI has revolutionized this field and demonstrated considerable potential for computer-assisted diagnosis. To the best of our knowledge, this is one of the first studies to develop a VIA-AI algorithm using low-power colposcopy images from local sources to match ethnicity and morphology. The newly developed algorithm was validated using a real-life screening and diagnosis process. This VIA-AI tool accurately detected AW features, ranging from 72% to 78% and 82% of invasive cancers. However, several investigators have demonstrated the feasibility of AI in colposcopy using various transfer-learning modalities. Simoes et al12 demonstrated a diagnostic accuracy of 72.2% and Miyagi et al13 demonstrated an accuracy of 82.3% in detecting high-grade lesions of the cervix on colposcopy. When piloted in a real-life program, this tool showed promising data with an accuracy of 76% compared with the gold standard test. This indicates that such technology can be successfully integrated as a key adjunct to a screening program, especially in low-resource settings, while maintaining high quality standards. A few previous studies showed that AI algorithm using cervicography images can automatically distinguish between normal and abnormal (CIN2+) cervical tissues after acetic acid application as accurately as 91%. However, most of these studies were performed in controlled settings with selected image data sets.9,10 This study compared the diagnosis of cervical precancers and cancers using the gold standard (histopathology) and reference standard diagnosis (colposcopy Swede score)-based categorization of VIA-screen–positive women requiring treatment in a real-world screening program conducted in a low-resource setting. The sensitivity and specificity of the developed AI algorithm are 62.5% and 97.6%, respectively. The AUC is 76.01 (Fig 4). This indicates that the AI tool developed using images collected from local sources could provide accurate guidance for nurses to decide on the need for treatment in the screen-and-treat screening programs during the same visit, even in the absence of direct supervision by experienced colposcopists. However, internal validation of an automated visual evaluation tool in Zambia identified cervical precancerous and cancerous (CIN2+) lesions with high-performance AUC = 0.91 (95% CI, 0.89 to 0.93), indicating a better sensitivity than VIA test alone.14 Nakisige et al also demonstrated a diagnostic accuracy of 0.84 (95% CI, 0.75 to 0.93) for an AI decision support system for VIA on the basis of the expert consensus opinion as the gold standard.15 Furthermore, it is known that AI performance decreases with external validation.16 ROC curve of diagnostic performance under the different gold standards when VIA screening test is positive. AI, artificial intelligence; ROC, receiver operating characteristic; VIA, visual inspection with acetic acid application. This tool facilitates improvements in the treatment uptake of screen-positive women, thereby reducing the number of visits to facilities in remote areas. This is one of the first studies to be conducted using an AI tool developed for decision-making guidance for nurses in a real-world resource-scarce situation. Testing this tool in different study populations and settings is necessary to fully understand its performance. The three major strengths of our study to develop and validate the VIA-AI tool performance were that this tool was developed from images collected from native women and was tested in an uncontrolled real-world low-resource setting, with images tested gathered from mid-level providers. Our model failed to correctly predict cervical lesions in approximately 20% of the images, which is a major challenge (Table 4). The high false-positive rate with this AI tool is due to excess mucus, blood on the cervical epithelium, poor-quality images, and photograph glare spots mimicking AW areas of the cervix, resulting in incorrect interpretation (Fig 3). Approximately one third of the misinterpretations were due to image-capturing errors such as specular reflections as glare spots, poor quality, and blurred images. Another third of the cases were incorrectly classified because of excess mucus and blood in the cervical tissues, which prevented proper visualization of the cervix. These identified challenges can be rectified in the future with appropriate training of health care personnel. This proof-of-principle investigation has several limitations. It is difficult to assess the exact accuracy of the developed VIA-AI tool, because the optimal sample size and adequacy required for the development and performance of AI models with fewer errors have not yet been established. The VIA-AI tool was developed on the basis of a highly selected data set of images corresponding to abnormal morphologies in the absence of image noise features. This resulted in misinterpretation and a high false-positive rate when the VIA-AI tool was used in a real-time program. Although the results are promising, further studies are required to upgrade and improve the AI algorithms. Our goal was to implement a user-friendly AI algorithm that can distinguish between cervical neoplastic lesions that require treatment with normal epithelium and benign lesions that does not require treatment. The prevalence of cervical inflammation and large ectropion of the cervix, a common feature in low socioeconomic regions, was also prevalent in our study regions, catering to women from rural backgrounds. This also resulted in incorrect interpretations by the VIA-AI tool, ranging up to 42%. On the basis of these observations, the VIA-AI tool can be revised and updated by including images with the abovesaid features. These false-negative features can be avoided by providing proper training to capture high-quality images. In conclusion, using histologic assessment as the gold standard, the developed VIA-AI tool achieved a sensitivity and specificity of 62.5% and, 97.6% respectively, and an ROC accuracy of 0.76 in a real-life setting in an LMIC, indicating that it is a promising technology for low-resource settings and provides scope for further improvements. This VIA-AI algorithm is a simple, rapid, noninvasive, easy-to-use, and cost-effective tool for guiding mid-level health care providers in rural and remote areas to provide quality assured care without the direct supervision of physicians as point-of-care clinical decisions, and would reduce unnecessary treatments and their associated financial burden. ACKNOWLEDGMENT The authors thank (1) Patricia D. Bidinger and team, Institute for Rural Health Studies, Hyderabad, India; and (2) B. Ravi, S. Srinivas, and C. Vijay, AGAMA Systems Developers, Hyderabad, India for their support. APPENDIX 1. PHASE I: DEVELOPMENT OF THE ARTIFICIAL INTELLIGENCE ALGORITHM VISUAL INSPECTION WITH ACETIC ACID APPLICATION-ARTIFICIAL INTELLIGENCE SOFTWARE The process for the development of the visual inspection with acetic acid application-artificial intelligence (VIA-AI) tool consisted of using low-power colposcopy images to visualize cervical changes that were highlighted with the application of exogenous contrast agents such as 3%-5% acetic acid. The system was developed by keeping this principle at the forefront and using cervical images of women from local settings to best represent morphologic changes in women in this region. The algorithm was developed so that it could identify and categorize the VIA-positive lesions into binary classification of normal—those that do not require any immediate treatment—or abnormal, that is, (1) those that could undergo immediate ablative treatment or (2) those to be referred for excisional methods, and (3) those with suspected cancers for definitive treatment procedures. Image Collection The archived data of low-power colposcopy images were retrieved. The data set consisted of the digital images collected from the binocular colposcopies and images captured from the video colposcope monitor from the women undergoing VIA-based cervical screening. Low-power colposcopy images after 4% acetic acid application with diagnostic confirmation by histopathologic examination of biopsies or those with colposcopy diagnosis confirmed by the expert colposcopists were included. Only white-light images with resolutions of 640 × 480 pixels were used. The colposcopy images were retrieved from around 2,600 women. All personal identification details were removed from these data. A total of 2,186 good-quality images were identified and each woman had only a single image. Images with poor illumination, substantial blur, lack of focus, insufficient visual capture of the cervix, and obstruction with instruments were discarded. All images were re-evaluated and categorized by an expert colposcopist. Image Classification and Descriptive Statistics of Images All images with acetowhite (AW) areas had histologic confirmation or colposcopy confirmation by modified Swede's score. There were 449 VIA-negative images and 867 were VIA-positive images. The positive images consisted of 329 were those with squamous metaplasia who do not need any treatment, 331 were of those who need ablative treatment, those who need a referral to a higher center for excisional treatment were 51 and 156 with invasive cancer (Fig 1). The images were broadly defined as normal (VIA-negative) when only light pink-colored, smooth epithelium corresponding to normal was observed without any AW areas (Fig 2A), and abnormal (VIA-positive) when AW areas are observed in the transformation zone (TZ) attached to the squamocolumnar junction (SCJ), which corresponds to abnormal lesions. The abnormalities (VIA-positive) were again subcategorized on the basis of the intensity of the color of the AW areas, surface and margins, and size of the lesions (Fig 2B), suspicious of invasive cancer—irregular, raised, ulceroproliferative growth involving the cervix or in the endocervical canal, bleeding points on touch, surface irregularity, mountain-like raised chalky white areas, and atypical vessels (Fig 2C). Categorization of VIA-Positive Images These were performed on the basis of the AW areas and their characteristics. AW area margins were scored as Score 0 indicating no margins, thin diffuse margins, thin hair-like flat line-like, ragged/feathery, thin irregular line-like, thin interrupted lines, indistinct, and cloudy-like merging edges, and Score 1 for AW areas with clear-cut, distinct, well-demarcated straight or round, raised margins, and cleft-like or ridge-like, mountain-like margins. The intensity of the AW area was scored as Score 0 indicating an absent AW area, thin, faint transparent white, glossy white, flat, cloudy white, fine mosaic, and thin crypt openings. A Score 1 indicated the AW area in the TZ, attached to the SCJ; dense, thick, raised, dull opaque, chalky white AW area, with or without cuffed crypt openings; and coarse mosaics. Labeling of AW Areas AW areas were labeled as follows: those that do not require any treatment (normal/benign), and those that require treatment (abnormal or precancers), which are included as (1) lesions that can be treated by the nurses using ablative methods as screen and treatment method (lesions confirmed as cervical intraepithelial neoplasia grade 1, 2, and 3 on the basis of the histopathologic confirmation or those confirmed by the modified Swede score of >4 on colposcopy). Lesions that are confined to the ectocervix are clearly visible, occupy 75% of the cervix, and do not extend inside the endocervical canal or onto vaginal fornices are indicated for ablative treatment. (2) Lesions of a higher grade and extensive size occupying the entire cervix extending inside the endocervical canal are indicated for referral to a higher center for excisional procedures. (3) Lesions with features of invasive changes require referral to a cancer center for definitive treatment. Image processing and preprocessing of data sets were performed successfully to categorize them as normal, abnormal, and suspicious invasive cancers. The data sets were constructed using images with very good accuracy. Data Augmentation For the training data set, data augmentation was performed for the less frequent classes to overcome data imbalance limitations. Hence, more advanced methods for image augmentation have been developed to improve accuracy. Multiple image augmentation policies were applied to increase the size and generalize the images. Generic augmentation techniques such as center cropping, horizontal flips, vertical flips, resizing, random resized crops, random rotation, and data set–specific augmentation techniques have been used to enhance the brightness of images, such as contrast-limited adaptive histogram equalization. Contrast- and brightness-enhancing filters were applied to enhance the local contrast of the images and to increase the visibility of the whiteness of the AW areas. This augmentation was customized for each training data set built for each model by adding images that were rotated by 30°, 60°, 90°, 120°, and 150° for horizontal, vertical, and horizontal-vertical flipping of the data sets. In addition to the above data augmentation techniques, machine learning image augmentation techniques such as RandAug and AutoAug were used to confirm the results (Appendix Table A2). Construction of Data Sets Training data sets and testing/validation data sets were randomly selected at the ratio of 80:20 from the total available images. The characteristics of the image data sets are presented in Appendix Table A1. For transfer learning of the convoluted neural network sets with internal validation, a combination of efficient V2 neural network models was used to obtain the final results and test these trained models on a new set of approximately 100 images for internal validation. Models were used to make inferences about these images and the results were compared (Appendix Table A2). Efficient Net B3 and V2s yielded the best accuracies compared with all the models tested. During the process of obtaining the best models on these data, Gradient-weighted Class Activation Mapping was used to visualize the activations of the trained models, which were correlated with the physician's observations in key areas such as the TZ and AW areas in the images. Training was performed using the transfer learning method, in which only the final layers of the data sets were trained by maintaining the weights of all the remaining frozen layers. Appendix Table A2 shows the output-based results for the models with the best final results. Prototype development and cross-validation of the VIA-AI tool was performed, and a real-time automated application was developed for use on mobile phones or tablet devices by nurses and health care workers at health centers. After the models were developed, this tool was tested on another separate testing data set consisting of 100 images for cross-validation. These images were not introduced into the software before testing. The main outcome measures and statistical analysis showed an AI prediction accuracy of 78% in identifying correct intensity scores of AW areas, 72% in identifying correct margin scores of AW areas, and 82% in identifying VIA-positives suspicious of invasive cancer. TABLE A1Distribution of Training Data SetsLesion CharacteristicSourceWithout Data AugmentationWith Data AugmentationTraining SetValidation SetTraining SetValidation SetAW area intensity Score 0, 1427342853,420850 Score 2400320803,200800AW area margins Score 0, 16515051265,0501,260 Score 25434301084,3001,080VIA-positives for treatment Ablative treatment317254632,540630 Referral to higher center339271682,710680Abbreviations: AW, acetowhite; VIA, visual inspection with acetic acid application. Distribution of Training Data Sets Abbreviations: AW, acetowhite; VIA, visual inspection with acetic acid application. TABLE A2Accuracy of Data Augmentation With Different Training Models StudiedTraining ModelWithout AugmentationWith AugmentationAW area intensity Efficient net B3_NS77% (18/25)78% (115/156) Efficientnet_lite268% (28/50)69% (37/50) Efficientnet_lite367% (27/50)72% (39/50) Efficientnet_lite463% (31/50)74% (34/50) Efficientnet_V2s70% (37/50)76% (34/50)AW area margins Efficientnet_B3_pruned74% (129/156) Efficientnet_B0_NS71% (128/156) Efficientnet_B1_NS72% (134/156) Efficientnet_B3_pruned68% (34/58)70% (43/58) Efficientnet_B0_NS66% (30/58)72% (43/58) Efficientnet_B1_NS65% (39/58)69% (43/58)VIA-positives for treatment Efficientnet_B3_pruined86% Efficientnet_B0_NS83% Efficientnet_B1_NS84%Abbreviations: AW, acetowhite; VIA, visual inspection with acetic acid application. Accuracy of Data Augmentation With Different Training Models Studied Abbreviations: AW, acetowhite; VIA, visual inspection with acetic acid application. DISCLAIMER Any opinions, findings, and conclusions expressed in this material are those of the author(s) and do not necessarily reflect those of ASCO or Conquer Cancer Foundation. SUPPORT Supported by the Conquer Cancer Foundation of the ASCO International Innovation Grant (2021). AUTHOR CONTRIBUTIONS Conception and design: Usha Rani Poli, Anirudh G. Gudlavalleti, G.V.S. Murthy Administrative support: Usha Rani Poli, G.V.S. Murthy Provision of study materials or patients: Usha Rani Poli, Anirudh G. Gudlavalleti, Jaya Bharadwaj Y Collection and assembly of data: Usha Rani Poli, Anirudh G. Gudlavalleti, Jaya Bharadwaj Y Data analysis and interpretation: Usha Rani Poli, Hira B. Pant, Varun Agiwal Manuscript writing: All authors Final approval of manuscript: All authors Accountable for all aspects of the work: All authors AUTHORS' DISCLOSURES OF POTENTIAL CONFLICTS OF INTEREST The following represents disclosure information provided by authors of this manuscript. All relationships are considered compensated unless otherwise noted. Relationships are self-held unless noted. I = Immediate Family Member, Inst = My Institution. Relationships may not relate to the subject matter of this manuscript. For more information about ASCO's conflict of interest policy, please refer to www.asco.org/rwc or ascopubs.org/go/authors/author-center. Open Payments is a public database containing information reported by companies about payments made to US-licensed physicians (Open Payments). No potential conflicts of interest were reported. REFERENCES