Comparing Visual Inspection with Acetic Acid to Pocket Colposcopy for Triage of HPV+ women living with HIV in Kisumu, Kenya

Introduction: : Cervical cancer is the fourth most common cancer among women worldwide, with nearly 600,000 new cases and 330,000 deaths each year. Women in low- and middle-income countries bear a disproportionate burden due to limited access to prevention tools. HPV testing is widely accepted as a screening tool for cervical cancer. However, the reduced specificity of HPV testing can cause overtreatment, strain healthcare systems, and miss cancers, prompting the WHO to recommend a triage step. While colposcopy is the gold-standard triage in high-income settings, it is expensive, requires trained personnel, and involves multiple visits—leading to loss to follow-up. In contrast, visual inspection with acetic acid (VIA), commonly used in low-resource settings, is more affordable but has high variability and low sensitivity due to poor lighting, lack of magnification, lesion size, and limited training.
To address these gaps, we developed the Pocket Colposcope (Pocket)—a low-cost, mobile colposcope powered by a phone and the Calla Health App for image and data storage (Figure 1a). Inserted through a speculum, the Pocket captures cervical images using widefield and high-resolution zoom modes. The Pocket, which captures images comparable in quality to a standard colposcope, offers an opportunity to enhance VIA through improved magnification and contrast. We conducted two clinical trials with HPV-positive women living with HIV (WLWH) in Kisumu, Kenya, in collaboration with the Kenya Medical Research Institute (KEMRI), to evaluate the potential of improving VIA in a real-world setting.

Materials and

Methods: : We performed two studies under IRB approvals from KEMRI and Duke (KEMRI #KEMRI/SERU/4318; Duke #Pro00106169). All participants provided informed consent.
Study 1 compared standard VIA (n=191) to Pocket colposcopy (n=197) in a randomized clinical trial (Figure 1b). Study 2 was a single-arm Pocket-only study (n=214). A clinical officer and Duke staff trained the KEMRI team, who then trained local providers using a “train-the-trainer” model. Biopsies were taken of suspicious lesions or two random areas if none were observed. Histopathology was categorized as NILM (no malignancy), cervical intraepithelial neoplasia (CIN) 1, 2, or 3. Participants received ablative treatment per WHO guidelines or were referred for advanced care. Endpoints included sensitivity, specificity, positive predictive value (PPV), and negative predictive value (NPV).
A feedback loop was implemented to improve image quality—particularly by reducing blur, a key factor for diagnostic accuracy. Image quality was assessed using an automated blur detection algorithm developed by our group, which categorized Pocket images as “Blurry” or “Clear” (Ekem et al. 2024). This algorithm includes object detection to delineate the cervix from non-relevant regions. FFT analysis was used for ground truth labeling, and a YOLOv5 object detection algorithm for blur classification and region of interest delineation. Duke staff met weekly with the KEMRI team to review algorithm-classified images and provided feedback on strategies for capturing focused images, such as stabilizing the Pocket on the speculum, slowly advancing toward the cervix to find the working distance and repositioning to reduce glare.

Results, Conclusions, and Discussions:: In study 1, Pocket and VIA performed similarly on sensitivity (26.3% vs 25.0%), specificity (88.9% vs 84.0%), and NPV (82.9% vs 87.1%). However, the PPV for the Pocket arm was nearly twice that of VIA (37.5% vs 20.6%), indicating it detected significantly more treatable disease (Table 1). Providers and participants completed surveys after each exam. Providers reported high satisfaction with the Pocket, noting minimal challenges, usability of the Calla Health App, and adequate internet access. Nearly all participants were confident in the provider’s assessment (Pocket: 96.8%, VIA: 97.4%) and would recommend the exam (Pocket: 98.9%, VIA: 99.5%). In study 2, the sensitivity of the Pocket more than doubled (55.6%), and NPV increased to 94.3%, indicating greater ability to detect disease and reassure when exams were negative. Specificity decreased to 70.6%, and PPV dropped to 15.4%, reflecting a shift toward more true positives at the expense of false positives—a trade-off often acceptable in screening contexts where minimizing missed disease is critical.
We hypothesized that the weekly feedback loop using the blur metrics would improve image quality and the number of participants with at least one clear image over time. Examples of algorithm annotated images are shown in Figure 2(a-d). As shown in Figure 2(e), the percent of participants with at least one image classified as “Clear” by the automated blur detection algorithm increased dramatically over the course of study 1 and modestly over the course of study 2 reflecting improvements in image acquisition and highlighting the importance of real-time feedback and image quality monitoring for maximizing performance.
These findings demonstrate that the Pocket—paired with targeted training and image quality support—can substantially improve CIN2/3 detection over VIA. Future integration of automated quality control could further enhance image quality and diagnostic accuracy (Ekem et al. 2024). Together, these studies support the Pocket colposcope as a feasible, lower-cost alternative to standard colposcopy, suitable for facilitating biopsy confirmation, provider training, patient education, and improved diagnosis.

Acknowledgements and/or References (Optional):: The team would like to acknowledge the significant efforts of the Kenya Medical Research Institute (KEMRI). Jeniffer Ambaka, study coordinator, Saduma Ibrahim, Data Manager, Evans Obuto and the three research assistants that supported data collection.