Cervical cancer is the second most common cause of cancer death in women worldwide,(1) with 85% of cases and cervical cancer deaths occurring in low-and middle-income countries (LMICs).(1–3) The decline in cervical cancer incidence and mortality in high-income countries since the 1970s is largely credited to effective cytology-based screening programs.(4,5) However, such screening programs require complex infrastructure and high quality health care services, (6,7) which most low- income countries cannot afford.(8) The current international guidelines advocate the use of human papillomavirus (HPV) testing for first line triage.(9) Self HPV testing, screening followed by Visual Inspection with Acetic acid (VIA) test improves specificity, minimizing over treatment inherent in one-visit screen- and-treat programs.(10) The diagnosis and excision of precancerous cervical lesions early in their development phase successfully prevents the formation of invasive cervical cancer.(11) There are ongoing international investigations of methods to automate current methods of cervical assessment (VIA, colposcopy) in order to increase accuracy of detection of high risk precancerous lesions, and streamline the intense monitoring and evaluation required for community health providers.(12) In low-resource nations, visual inspection of the cervix with acetic acid (VIA) is now the most used technique for cervical cancer screening test.(13) Screening with the VIA test is regarded as the most viable and practical technique for early diagnosis of cervical cancer, as well as the most beneficial in terms of, for instance, lowering disease burden at a reasonable cost.(14) Recent research has shown visual inspection with acetic acid (VIA) as an alternative sensitive screening approach.(15) It is inexpensive and non-invasive, and it may be performed at a low-level health institution such as a health center. More significantly, VIA gives fast results, and individuals who are suitable for treatment can undergo cryotherapy treatment for precancerous lesions on the same day and in the same health institution.(16) National guideline for the prevention and screening of cervical cancer in Nepal has emphasized using the VIA approach for screening.(17) While the government has pledged to eliminate cervical cancer through the adoption of a good screening program, there are still obstacles to the policy’s implementation.(14) The lack of well- trained medical human resources, as well as the subjective nature of the test, have added hurdles to VIA’s broad reach. Visual examination procedures have been shown to be subjective, and diagnostic performance varies significantly, with physicians attaining better results than nurses and paramedics.(18) Unfortunately, competent physicians who can appropriately interpret VIA may not be easily available in developing countries like Nepal. (14,19,20) Digital cervicography, or digital image of the cervix following acetic acid application, has proved extremely useful in increasing quality control and is also considered as an efficient technique for eliminating interpreters’ subjectivity. Recent advancements in smartphone technology have opened up new opportunities for cervical screening in low-resource settings, addressing the constraints of colposcopy, such as device`s bulkiness, high cost, electricity dependency, and ongoing maintenance requirements.(21–23) Machine Learning (ML), in particular, Deep learning (DL) which is a type of ML approach, has shown promising results in medical image-based diagnosis. Some of the DL based tools have already started being commercially deployed in real clinical settings. DL has recently been explored in classifying cervix images for VIA screening with cervix image, and some of these works have shown the feasibility of DL for VIA with encouraging results.(22-25) As a result, in this project, we propose developing and evaluating an ML-based VIA screening tool using data collected from Nepali women population.

General Objective:

• To develop and evaluate an ML-based VIA screening tool using data collected from Nepali women population.

Specific Objective:

• Create an annotated cervix image (indicating cervical cancer regions) dataset after using acetic acid using mobile phone captured and medical grade colposcopy.
• Develop prototype of ML-based cervical cancer screening tool based on using VIA that uses cameras housed within mobile phones.
• Evaluate and perform a validation study of the ML-based cervical cancer screening tool in real-world settings.

Expected Impact

Impact on primary beneficiary group:

 The project has a huge impact on the health and quality of lives of women in low-resource settings. The direct beneficiaries will be the 30-60 years old women who have to undergo cervical cancer screening at least every three years according to the national cervical cancer screening guideline in Nepal.(26) Currently, the screening coverage of cervical cancer is 8.2%.(27) In order to eliminate cervical cancer, screening coverage has to be at least 70%. This technology will make it possible to reach women in rural areas with limited infrastructure and human resources. Over the last three decades, the widespread use of screening in high-income nations has resulted in a remarkable decrease in cervical cancer deaths. According to the International Agency for Research on Cancer (IARC), there is substantial evidence that cervical cancer screening can reduce deaths due to cervical cancer by 80 percent or more among screened women.(28) Additionally, other direct beneficiaries will be health care providers, whose service delivery will be significantly improved with the application of this technology.

Technological impact:

VIA is a commonly utilized method for cervical cancer screening in places because of its ease where more advanced screening procedures are not accessible. However, the interpretation of the results from VIA is subjective, and hence not considered reliable. This project will aid in the development of an automated visual evaluation system that is simple to use, that will enable to standardize the result of the VIA. The ML model developed in this project will perform better than all standard screening tests at predicting case identification. The ability of the ML model to provide automated visual evaluation for VIA screening on par or higher than a human expert will revolutionize scaling of the screening in LMICs where lack of expertise in most parts prevents routine screening of a large fraction of the population. Moreover, in urban centers it also allows already overburdened doctors to deliver services in less time and low cost. The success of this project will not only advance cervical cancer screening technology but will also provide a great example of locally led multidisciplinary project building data infrastructure, technology stack for managing large imaging data, and deploying ML models in a web-based platform. This will open up similar applications where ML in medical imaging data has potential.

Build research capacity:

This project will build the capacity of the researchers to convert the evidence into commercial products. The project will encourage researchers and the team to advance national, regional and international networks. This project has brought interdisciplinary teams together from several sectors involved in health research projects, including investigators, data managers, software developers, data scientists, public health practitioners, administrators, and institutional authorities. Development of skills in project planning and cervical cancer is 8.2%.(27) In order to eliminate cervical cancer, screening coverage has to be at least 70%. This technology will make it possible to reach women in rural areas with limited infrastructure and human resources. Over the last three decades, the widespread use of screening in high-income nations has resulted in a remarkable decrease in cervical cancer deaths. According to the International Agency for Research on Cancer (IARC), there is substantial evidence that cervical cancer screening can reduce deaths due to cervical cancer by 80 percent or more among screened women.(28) Additionally, other direct beneficiaries will be health care providers, whose service delivery will be significantly improved with the application of this technology.