An Automated Multimodal, Multiscale Algorithm for Cervical Cancer Screening in Low- and Middle-Income Countries

PROJECT SUMMARY/ABSTRACT Cervical cancer is responsible for the deaths of over 300,000 women per year. An estimated 90% of the deaths occur in low- and middle-income countries (LMICs). Due to the limited access to healthcare facilities in LMICs and time required for staining a specimen and receiving pathology confirmation of a diagnosis, the WHO recommends screening with an HPV DNA test followed by triage and further treatment in a single visit. Multimodal cervical imaging integrated with deep learning algorithms can serve as an automated, reproducible, and scalable triage test that provides a real-time highly sensitive and highly specific diagnostic result in LMICs, where there may be a lack of trained clinicians and pathologists. The Richards-Kortum lab has developed a low-cost multimodal colposcope that captures microscopic and macroscopic images of the cervix in a clinical setting. Building on these hardware developments, the deep learning-based diagnostic test will be designed by pursuing the following three specific aims: 1) Design a multimodal image registration algorithm to spatially correlate and combine information from widefield colposcope images of the cervix, that are acquired after staining with each of two common standard of care solutions (acetic acid and Lugol’s iodine), and HRME images. 2) Develop a deep learning algorithm that integrates spatiotemporal information across a sequence of HRME video frames and predicts cervical cancer diagnosis in real-time. 3) Evaluate clinical performance of the algorithms using an available dataset of images collected from patients at two different low-resource setting clinical sites in Brazil. This project benefits greatly from the co-mentorship offered by Dr. Rebecca Richards-Kortum and her colleague Dr. Kathleen Schmeler, who both have extensive experience in cervical cancer prevention for low-resource settings. Dr. Richards-Kortum has led many projects involving the development of low-cost biomedical optical imaging technologies and their associated algorithms. Dr. Schmeler has led efforts to enable integration of these technologies into clinical workflows. The training plan and environment at Rice University and Baylor College of Medicine have been specifically tailored for my interests in deep learning, biomedical optical imaging, clinical translation, and global health. My training involves receiving guidance and feedback from experts in deep learning for biomedical applications as I develop the proposed algorithms, attending workshops, presenting at imaging and global health conferences, teaching and mentoring trainees, and pursuing translational research aims to ensure success in my proposed research. Overall, my proposed research and training plan aligns perfectly with my goal of becoming a physician-scientist in the field of obstetrics & gynecology. As a physician- scientist, I hope to continue working on the development of biomedical imaging technologies for cancer screening and facilitate the integration of these technologies into the clinical care I provide for underprivileged communities.