Applying AI models to digital placental photographs to automate and improve morphology assessments

This study introduces PlacentaVision, an AI-based model that automates placental morphology measurements from digital photographs with high accuracy and standardization compared to human assessments, while highlighting that discrepancies are more pronounced in irregularly shaped or preterm placentas.

The Gist

Imagine the placenta as the ultimate life-support system for a baby growing in the womb. It's like a high-tech power plant and food delivery service rolled into one. For decades, doctors have known that the size and shape of this "power plant" can tell us a lot about a baby's future health, from their heart health to their risk of high blood pressure later in life.

However, measuring this organ after birth has been a bit of a mess. It's like trying to measure a squishy, irregularly shaped potato with a ruler while wearing oven mitts. Different doctors measure it differently, they make mistakes, and often, they just don't measure it at all because it's too much work.

Enter "PlacentaVision": The AI Super-Observer

This paper introduces a new tool called PlacentaVision. Think of it as a super-smart robot camera that looks at a photo of a placenta and instantly measures it with perfect consistency. The researchers wanted to see if this robot could do a better job than the human pathologists who usually do the measuring.

Here is the breakdown of their adventure, explained simply:

1. The Mission: Robot vs. Human

The team took photos of over 27,000 placentas from three different hospitals: two in the US (Chicago and Pittsburgh) and one in Uganda.

• The Humans: Trained technicians measured the placentas by hand using a physical ruler and wrote down the numbers.
• The Robot (AI): The PlacentaVision software looked at the digital photos, found a ruler in the picture, and calculated the length and width automatically.

They then compared the two sets of numbers to see who was closer to the "truth."

2. The Results: A Close Race

The results were surprisingly good!

• The Average Gap: On average, the robot's measurements were less than half an inch (about 0.6 cm) different from the human measurements. That's roughly the width of a standard pencil eraser.
• The "Good" News: About half the time, the robot and the human were within one centimeter of each other.
• The "Bad" News: Sometimes the gap was bigger. The robot and the human disagreed the most when:
  • The placenta was weirdly shaped (like a potato with a bump or a lopsided pancake).
  • The baby was born prematurely (smaller placentas are harder to measure).
  • The photo was taken at the Chicago hospital (likely due to different lighting or photo styles).

3. Why the Disagreement? (The "Why" Behind the Numbers)

The researchers dug into why they disagreed, and it wasn't just the robot making mistakes.

• The "Potato" Problem: When a placenta is round and flat, it's easy to measure. But when it's lumpy or irregular, humans struggle to decide exactly where the "edge" is. The robot, however, has a strict mathematical rule for finding the edge, so it's consistent, even if humans find it confusing.
• The "Typo" Factor: Sometimes the human measurements were just wrong because of data entry errors. Imagine a placenta is 20 cm long, but the technician accidentally types "30 cm" into the computer. The robot, looking at the photo, correctly says "20 cm." In these cases, the robot was actually the hero!
• The "Membrane" Confusion: Sometimes the robot got confused. If the thin, transparent membranes surrounding the placenta were in the photo, the robot might think they were part of the main disc and measure them too, making the placenta look bigger than it is.

4. The Big Picture: Why This Matters

Think of the placenta as a fossil record of pregnancy. If we can measure it perfectly and automatically, we can unlock a treasure trove of health data.

• Standardization: Right now, one doctor's "long" might be another doctor's "short." PlacentaVision gives everyone the exact same ruler and the exact same rules.
• Scale: Humans can only measure a few placentas a day. An AI can measure thousands in the time it takes to drink a coffee. This means we could eventually check every placenta, not just the ones that look suspicious.
• Future Health: By automating this, we can build a massive database linking placenta shapes to future diseases, helping doctors predict and prevent health issues for children before they even happen.

The Bottom Line

The paper concludes that AI is ready to be the new standard for measuring placentas. It's not perfect yet (it still gets confused by weird shapes and bad photos), but it is incredibly close to human accuracy and far more consistent.

In short: The robot is learning to be a better measurer than the human, and it's going to help us understand how babies grow and stay healthy for a lifetime.

Technical Summary

1. Problem Statement

The human placenta is a critical indicator of fetal development and long-term offspring health outcomes (e.g., cardiovascular disease, hypertension, and mortality). However, current clinical assessment of placental morphology is hindered by:

• Lack of Standardization: Protocols for measuring placental dimensions (length, width, shape) vary across hospitals and laboratories.
• Human Error and Variability: Manual gross pathology examinations are prone to inter-observer variability, data entry errors, and "digit preference" (rounding numbers).
• Resource Constraints: Only ~20% of placentas in the US undergo pathological examination due to the high resource burden and shortage of pathologists.
• Inconsistency in Irregular Cases: Measuring irregularly shaped placentas (e.g., bilobed or fragmented) is particularly difficult for humans to standardize.

The study aims to address these issues by developing and validating PlacentaVision, an AI-driven system that automatically measures placental morphology from digital photographs, comparing its accuracy against human gross pathology measurements.

2. Methodology

Study Design and Data Collection

• Type: Multi-site comparative study.
• Sites: Three hospitals with diverse demographics and resource settings:
  1. Northwestern Memorial Hospital (Chicago, USA): Tertiary care; data extracted from electronic medical records (EMR) of standard pathology workflows.
  2. UPMC Magee-Womens Hospital (Pittsburgh, USA): Tertiary care; data from the MOMI Biobank research study.
  3. Mbarara Regional Referral Hospital (Uganda): Low-resource setting; data from prospective research studies (PACO and PlacentaVision).
• Sample Size: Initial dataset of 38,336 singleton births. After exclusions (missing images, poor quality, fragmented discs, missing rulers, implausible measurements), the final analytic sample was 27,846 placentas.
• Imaging Protocol: Placentas were photographed on a blue background with a centimeter ruler. The protocol specified positioning for fetal and maternal sides. Images were captured using various digital cameras (DSLRs and point-and-shoot) under varying lighting conditions.

The PlacentaVision AI System

The system utilizes a two-stage deep learning and computer vision pipeline:

1. Semantic Segmentation: The model segments the placental disc, identifies the umbilical cord insertion point, and classifies the image side (fetal vs. maternal).
2. Morphological Calculation:
   • Ruler Reading: The model detects the ruler in the image to establish a scale.
   • Dimension Extraction: It calculates the Long Axis (maximal linear dimension) and Short Axis (greatest dimension perpendicular to the long axis) following Amsterdam consensus criteria.
   • Shape Analysis: It overlays a best-fit ellipse to calculate a Shape Ratio (area inside/outside ellipse vs. ellipse area). A ratio >0.1 is classified as "irregular."
   • Additional Metrics: The AI calculates disk area, perimeter, and estimated weight, metrics often not captured in standard gross exams.

Statistical Analysis

• Comparison: Differences were calculated as $PlacentaVision - Human\ Gross\ Exam$.
• Metrics:
  • Bland-Altman Analysis: To determine limits of agreement (LoA) and bias.
  • Lin's Concordance Correlation Coefficient (CCC): To assess agreement relative to the 45-degree line of perfect agreement.
  • Stratification: Analyses were stratified by site, disc shape (regular vs. irregular), infant sex, and gestational age (preterm vs. term).

3. Key Results

Overall Agreement

• Mean Differences:
  • Length: AI measured 0.57 cm longer than humans (SD 2.19 cm).
  • Width: AI measured 0.25 cm longer than humans (SD 1.85 cm).
• Correlation:
  • Length CCC: 0.73 ($p<0.01$).
  • Width CCC: 0.69 ($p<0.01$).
• Limits of Agreement (Bland-Altman):
  • Length: -3.7 cm to +4.9 cm.
  • Width: -3.4 cm to +3.9 cm.
• Precision: Approximately 45% of length measurements and 50% of width measurements were within 1 cm of each other.

Stratified Findings

• Shape: Irregularly shaped placentas showed significantly larger discrepancies (Length difference: 1.53 cm) compared to regular/round shapes (0.45 cm). This suggests humans struggle more with defining axes on irregular discs.
• Site Variation:
  • Magee: Showed the closest agreement (mean difference ~0 cm).
  • Northwestern: Showed the largest mean difference (0.6 cm).
  • MUST (Uganda): Showed lower correlation coefficients (<0.55), likely due to image quality variations and different photography protocols.
• Gestational Age: Preterm placentas had slightly larger measurement differences (0.71 cm) compared to term placentas (0.53 cm).
• Sex: No significant difference in measurement agreement between male and female placentas.

Error Analysis

• AI Errors: Primarily caused by incorrect ruler reading (glare, poor contrast) or misclassifying membranes as part of the disc.
• Human Errors: Often attributed to data entry mistakes (e.g., transposing digits, recording 30 cm for a 20 cm disc) or inconsistent ruler placement.

4. Key Contributions

1. Validation of AI in Pathology: Demonstrated that AI can automate placental morphological assessment with high accuracy (sub-centimeter difference) compared to human experts.
2. Standardization: Provided a consistent, protocol-driven method for measuring placental dimensions, reducing inter-observer variability, especially for irregular shapes.
3. Scalability: Proved the feasibility of processing a massive dataset (n=27,846) across diverse global settings, a task impractical for manual pathology review.
4. New Metrics: Enabled the calculation of derived metrics (area, perimeter, shape ratio) that are not standard in routine gross exams but are valuable for research.
5. Identification of Human Bias: Highlighted that human measurements suffer from "digit preference" and significant inconsistency with irregular shapes, validating the need for automation.

5. Significance and Future Directions

• Clinical Impact: PlacentaVision offers a low-cost, scalable tool to standardize placental assessment, potentially allowing for the examination of all placentas rather than just the 20% currently selected for pathology. This could improve the detection of adverse pregnancy outcomes and inform long-term health research (DOHaD hypothesis).
• Research Utility: The ability to generate standardized, high-volume morphological data will enhance large-scale epidemiological studies linking placental structure to chronic diseases.
• Limitations & Future Work:
  • Current models struggle with low-quality images (glare, ruler visibility) and complex membrane/cord configurations.
  • Future iterations need to improve ruler detection algorithms and membrane segmentation.
  • Further research is needed to understand why human measurements deviate (e.g., via video recording of the measurement process) and to validate AI tools in clinical workflows.

Conclusion: The study confirms that AI-based PlacentaVision is a viable, accurate, and superior alternative for standardizing placental morphology assessments, particularly for irregular shapes and large-scale data collection, bridging the gap between limited pathology resources and the need for comprehensive perinatal data.