TL;DR
This systematic review explores the potential of AI algorithms for cervical cancer screening using images from Visual Inspection with Acetic Acid (VIA). While AI shows promise for objective and accessible screening, current studies often use small, highly selected datasets, limiting generalizability and requiring large-scale, real-world testing for clinical integration.
What Problem Does This Paper Address?
Cervical cancer remains a significant global health issue, particularly in low- and medium-income countries (LMICs). The World Health Organization (WHO) recommends early detection through screening. Visual Inspection with Acetic Acid (VIA) is a simple, low-cost method, but its effectiveness is hampered by high subjectivity and variability among healthcare providers. While histopathology is the gold standard for diagnosis, it's not always feasible in resource-limited settings. This paper addresses the challenge of improving the objectivity and accessibility of VIA-based cervical cancer screening by systematically reviewing the use of Artificial Intelligence (AI) algorithms to analyze VIA images.
Methods
The authors conducted a systematic literature search across PubMed, Google Scholar, and Scopus for studies published between January 2015 and July 2022. They focused on identifying automated AI algorithms designed to classify VIA images as either negative (healthy/benign) or positive (precancerous/cancerous). Inclusion criteria were strict, requiring algorithms to use only VIA images, have histopathology as the gold standard for positive cases, and be original, peer-reviewed research. For each included study, the algorithm with the highest reported accuracy was selected for analysis. The quality of the included studies was assessed using the QUADAS-2 tool. Data analysis involved comparing algorithms based on sensitivity and specificity, and qualitative analysis of preprocessing steps and classification techniques.
Datasets
The studies reviewed utilized various datasets, with a significant portion drawing from the "Guanacaste dataset," a collection of HPV and cervical cancer-related data gathered over seven years in Costa Rica. This dataset was often used in conjunction with a fixed-focus, ring-lit film camera called a cerviscope. Some studies also utilized data collected with smartphones, including the MobileODT EVA system. A notable aspect is that many studies used small, highly selected subsets of these datasets, which is a key limitation discussed later. No new benchmarks were introduced in this review; instead, it synthesized findings from existing studies.
Key Insights and Ideas
- AI's Potential for Objectivity: VIA's subjectivity is a major hurdle. AI offers a path towards more objective and consistent interpretation of VIA images, which is crucial for widespread screening.
- Accessibility through Portable Devices: The review highlights that AI algorithms performing well on images captured by smartphones and simpler cameras suggest a future where advanced screening tools can be more readily deployed in LMICs, bypassing the need for expensive colposcopes.
- The "Dynamic" Aspect of VIA: Some promising results emerged from algorithms that analyze multiple sequential images of the cervix over time, capturing the dynamic acetowhitening effect. This mirrors how human colposcopists assess lesions and suggests a richer source of information than single static images.
- AI Outperforms Human Experts (in some cases): In the few studies that directly compared AI performance to human experts, AI algorithms demonstrated higher sensitivity, and in some instances, better overall performance, underscoring their potential to augment or even surpass human interpretation.
Key Results
Out of 2608 identified studies, 11 met the inclusion criteria. The reviewed AI algorithms demonstrated a wide range of performance, with reported sensitivities and specificities ranging from 0.22 to 0.93 and 0.67 to 0.95, respectively. The study identified that algorithms using smartphones or cameras often achieved comparable or better performance than those using colposcopes. While AI algorithms generally outperformed human experts in sensitivity, some experts still showed higher specificity. The quality assessment revealed that most studies had a high risk of bias related to patient selection due to the use of small, highly selected datasets.
Limitations and Future Work
The authors acknowledge several limitations. The primary concern is the small dataset size and high selection bias in most included studies. This means the algorithms may not generalize well to real-world, diverse populations. The heterogeneity in performance reporting also made a direct meta-analysis impossible. Future work should focus on training and testing AI algorithms on larger, more diverse datasets that better reflect actual screened populations. Furthermore, standardization of preprocessing steps and image quality assessment is needed. Crucially, algorithms must be tested in real-world conditions using various devices and settings to ensure their feasibility and effectiveness for widespread cervical cancer screening, especially in LMICs.
Summary updated