A newly developed tool that leverages computer vision and artificial intelligence (AI) may help clinicians quickly assess placentas at birth, potentially improving neonatal and maternal care, according to new research from scientists at Northwestern Medicine and Penn State.
The study, which was published Dec. 13 in the print edition of the journal Patterns and featured on the magazine’s cover, describes a computer program called PlacentaVision that can analyze a simple photograph of the placenta to detect abnormalities associated with infection and neonatal sepsis, a life-threatening condition that affects millions of newborns worldwide.
The placenta is one of the most common specimens we see in the laboratory. When the neonatal intensive care unit is treating a sick child, even a few minutes can make all the difference in medical decision-making. With a diagnosis from these pictures, we can have an answer days earlier than we would in our normal process.”
Dr. Jeffery Goldstein, study co-author, director of perinatal pathology and associate professor of pathology at Northwestern University Feinberg School of Medicine
Northwestern provided the largest set of images for the study, and Goldstein led the development and troubleshooting of the algorithms.
Alison D. Gernand, the principal contact researcher for the project, conceived the initial idea for this tool through her work in global health, particularly with pregnancies where women give birth at home due to a lack of health care resources.
“Unexamined placental rejection is a common but often overlooked problem,” said Gernand, associate professor in the Department of Nutritional Sciences at Penn State’s College of Health and Human Development (HHD). “It’s a missed opportunity for identifying concerns and early intervention that can reduce complications and improve outcomes for both mother and baby.”
Why early placenta examination is important
The placenta plays a vital role in the health of both the pregnant woman and the baby during pregnancy, yet is often not thoroughly examined at birth, especially in areas with limited medical resources.
“This research could save lives and improve health outcomes,” said Yimu Pan, a doctoral candidate in the computer science program from the College of Information Sciences and Technology (IST) and lead author of the study. “It could make placenta screening more accessible, benefiting research and care for future pregnancies, especially for mothers and babies at higher risk of complications.”
Early recognition of placental infection through tools like PlacentaVision could allow clinicians to take early action, such as giving antibiotics to the mother or baby and closely monitoring the newborn for signs of infection, the scientists said.
PlacentaVision is intended for use in a range of medical demographics, according to the researchers.
“In low-resource areas — places where hospitals don’t have pathology labs or specialists — this tool could help doctors quickly identify issues like placental infections,” Pan said. “In well-equipped hospitals, the tool can ultimately help doctors identify which placentas need further, detailed examination, making the process more efficient and ensuring the most important cases are prioritized.”
“Before such a tool could be developed globally, the key technical hurdles we faced were making the model flexible enough to handle various placenta-related diagnoses and ensuring that the tool could be robust enough to handle various delivery conditions , including variation in lighting conditions, imaging quality and clinical settings,” said James Z. Wang, Distinguished Professor in the College of IST at Penn State and one of the principal investigators on the study. “Our AI tool needs to maintain accuracy even when many training images come from a well-equipped urban hospital. Ensuring that PlacentaVision can handle a wide range of real-world conditions was essential.”
How the tool learned how to analyze placenta images
The researchers used cross-modal adversarial learning, an AI method to align and understand the relationship between different types of data -. in this case, visual (images) and text (pathology reports) -? to teach a computer program how to analyze images of placentas. They gathered a large, diverse dataset of placental images and pathology reports over a 12-year period, studied how those images related to health outcomes, and built a model that could make predictions based on new images. The team also developed various image-shifting strategies to simulate different photo-taking conditions so that the robustness of the model could be properly evaluated.
The result was PlacentaCLIP+, a powerful machine learning model that can analyze placenta photos to detect health risks with high accuracy. Nationally validated to confirm consistent performance across populations.
According to the researchers, PlacentaVision is designed to be easy to use, potentially working through a smartphone app or integrated into medical record software so doctors can get quick answers after delivery.
Next step: A user-friendly application for medical personnel
“Our next steps include developing a user-friendly mobile app that can be used by medical professionals – with minimal training – in low-resource clinics or hospitals,” Pan said. “The user-friendly app will allow doctors and nurses to photograph placentas and receive immediate feedback and improve care.”
The researchers plan to make the tool even smarter by including more types of placental characteristics and adding clinical data to improve predictions while contributing to long-term health research. They will also test the tool in different hospitals to ensure it works in different settings.
“This tool has the potential to transform the way placentas are examined after birth, especially in parts of the world where these examinations are rarely done,” Gernand said. “This innovation promises greater accessibility in low- and high-resource settings. With further refinement, it has the potential to transform neonatal and maternal care by enabling timely, personalized interventions that prevent serious health outcomes and improve the lives of mothers and infants worldwide.”
This research was supported by the National Institutes of Health National Institute of Biomedical Imaging and Bioengineering (grant R01EB030130). The team used supercomputing resources from the Advanced Cyberinfrastructure Coordination Ecosystem: Services & Support (ACCESS) program funded by the National Science Foundation.
Source:
Journal Reference:
Pan, Y., et al. (2024). Cross-modal contrast learning for unified placental analysis using photographs. Patterns. doi.org/10.1016/j.pattern.2024.101097.