Bridging the Digital Divide in Cervical Cancer Detection: Integrating Deep Learning Diagnostic Algorithms into Low-Resource Clinics

Abstract

Cervical cancer accounts for over 660,000 new cases and 350,000 deaths annually, with more than 90% of deaths occurring in low- and middle-income countries (LMICs). Although the progression from initial HPV infection to invasive cancer typically spans 15–20 years, providing ample opportunity for prevention, barriers such as limited laboratory infrastructure, few trained providers, and low screening coverage prevent early detection and treatment. To expand access to screening, there has been a recent influx in affordable colposcopy imaging devices for providers to illuminate and magnify the cervix in search of visual cues of precancer; however, visual diagnosis suffers from poor specificity. To close these gaps, artificial intelligence (AI) methods can increase diagnostic consistency, reduce dependence on specialized expertise, and make image-based screening more scalable across diverse healthcare settings.Aim 1 develops and evaluates deep learning algorithms for automated classification of colposcopy images, moving beyond traditional machine learning approaches. A dataset of 1,180 patient visits across six countries was used, with paired acetic acid and green-light Pocket colposcopy images. Convolutional neural networks (CNNs) based on a ResNet-18 backbone were trained with preprocessing steps including automated cervix detection, specular reflection removal, and a class-balanced loss function. Importantly, multi-contrast image pairs (acetic acid and green light) were processed using both an image-stack and a parallel feature extraction method. The YOLOv3 cervix detection model resulted in an average precision of 0.997 on a test set of 83 acetic acid images. The automated detection model was used for the removal of the speculum and vaginal walls throughout all images prior to the training and test procedures. The best-performing model, consisting of parallel white and green light feature extraction, achieved an AUC of 0.87 with a sensitivity of 0.75, significantly outperforming single-contrast approaches (AUC 0.78–0.81, sensitivity 0.30–0.50). These results demonstrate that deep learning with multi-contrast inputs can substantially improve diagnostic performance of cervical cancer machine learning algorithms. Aim 2 addresses the challenge of adapting diagnostic algorithms to new devices, where small datasets and device-specific imaging characteristics hinder generalization. Using a domain-adaptive generative model, images from a standard-of-care (SOC) colposcope were translated into Pocket colposcope-like images while preserving anatomical structures and pathology labels. This approach augmented the limited Pocket dataset with diverse, high-quality synthetic samples. Generative Adversarial Networks were employed to align domains, expand training data, and mitigate issues of sampling imbalance. Compared to the original SOC, the synthesized Pocket images show improved alignment with the original Pocket images, with the Fréchet Inception Distance (FID) decreasing from 86.28 to 61.67 (a 38% reduction) and the Kernel Inception Distance (KID) decreasing from 0.067 to 0.048 (a 29% reduction). These results show that the synthesized Pocket images contain greater domain similarity to the original Pocket images compared to their SOC counterparts. When training a binary cervical cancer detection model using a Vision Transformer (ViT) architecture, incorporating generated Pocket-like images improved the model AUC from 0.698 (SOC pretraining alone) to above 0.77 after augmentation, with notable reductions in extreme misclassifications. We further validate our results across multiple model architectures and evaluate robustness to real-world factors such as reduced image quality and new clinical sites. By enhancing diagnostic performance on small, diverse datasets from accessible medical devices, this approach supports the expansion of cervical cancer screening in low-resource settings with the greatest need.Aim 3 builds on these advances by introducing a developer-to-user translational AI framework that addresses persistent challenges in implementing diagnostic algorithms into clinical settings. Because many algorithms are trained from carefully curated datasets, they often fail during clinical use due to large variabilities in image capture. The Calla Health mobile application (CHA) was specifically designed to make it easier for providers to collect more standardized and rich data—images and annotations—at the point of patient care to avoid error propagations that occur when training algorithms on real-world data. Through deployment testing of CHA, and its accompanying AWS server that hosts an image quality algorithm and a clinical database, providers in Kenyan clinics were able to capture clear colposcopy images and precise biopsy annotations with minimal disruptions to clinical workflow. Rather than treating model development as a static process, our platform bridges the gap between developers and clinical providers and supports continuous algorithm adaptation as new data is collected, ensuring that diagnostic tools remain relevant across populations. By pairing algorithmic innovation with deployment-focused infrastructure, this aim establishes a scalable pathway for translating AI diagnostics into routine screening programs, particularly in low-resource environments where sustainable impact is most needed.In summary, this dissertation presents a multi-layered approach to tackle the barriers facing the implementation of automated cervical cancer diagnostic models in low-resource settings. By tackling data scarcity through image contrast-combination techniques and domain adaptive generative models, we can improve diagnostic performance on newer, more accessible colposcopy devices, such as the Pocket. Furthermore, by integrating the Pocket and its diagnostic models into a single mobile application, the proposed framework demonstrates a practical pathway for real-world deployment. Ultimately, these methods support the integration of automated diagnostics into global screening efforts, promoting equitable access to the early detection of precancerous lesions and progress toward the World Health Organization’s cervical cancer elimination goals.