Dried Blood Spot Anemia Detection with CNNs

ISEF Category: Translational Medical Science

Subcategory: Disease Detection and Diagnosis  ·  Difficulty: Advanced  ·  Setup: School Lab  ·  Time: Full Year

The Hook

A tiny drop of blood can leave a ring pattern that looks random, but it can hide useful medical clues. If you can teach a computer to read that pattern, you may predict hemoglobin level without a full lab analyzer. That makes this project part imaging, part biology, and part machine learning. It also connects to a real need, low-cost anemia screening.

What Is It?

When a drop of blood dries on filter paper, the liquid does not spread evenly. Cells, proteins, salts, and water move at different rates. That can create a dark edge and a lighter center, often called a coffee-ring pattern. The final image can change with blood chemistry, including salinity and hematocrit, which is the fraction of blood made of red cells.

You can think of the dried spot like a fingerprint left by the blood as it dries. A camera sees the whole pattern at once, but a human cannot easily turn that picture into a hemoglobin estimate. A convolutional neural network, or CNN, is a type of machine learning model that learns visual patterns from images. In this project, the CNN tries to connect spot shape and texture to published reference values for hemoglobin and anemia status.

Why This Is a Good Topic

This is a strong science fair topic because you can test clear variables, measure image features, and compare predictions against a known target. It connects to anemia screening, which matters for public health, especially where lab testing is hard to access. You can also learn image analysis, model training, validation, and error checking, all skills that matter in real research.

Research Questions

• How does hematocrit change the shape and darkness of dried-blood-spot coffee-ring patterns?
• What is the effect of salinity on the ring width, edge sharpness, and center intensity of dried blood spots?
• Does a CNN trained on dried-blood-spot images predict hemoglobin better than a simple hand-measured image feature model?
• To what extent do published reference curves improve the calibration of hemoglobin predictions from spot images?
• Which image region, the center, the ring edge, or the full spot, gives the best anemia classification accuracy?
• How does paper type affect the relationship between dried-blood-spot pattern features and hemoglobin estimates?

Basic Materials

• USB microscope with adjustable focus and fixed stand.
• White filter paper or blood spot cards with known paper type.
• Safe blood substitute or de-identified dried-blood-spot image set from published sources.
• Smartphone or laptop camera for image capture.
• Ruler or printed scale card for image calibration.
• Consistent light source, such as a ring light or desk lamp with diffuser.
• Computer with enough storage for image files.
• Spreadsheet software for labeling and organizing samples.

Advanced Materials

• Access to de-identified dried-blood-spot samples with known hemoglobin values.
• Hemoglobin reference dataset from published studies.
• Controlled humidity storage container for drying consistency studies.
• Color calibration card for image standardization.
• High-resolution microscope camera or USB microscope.
• Access to a validated lab hemoglobin analyzer for comparison.
• Annotated image software for drawing spot boundaries and regions of interest.
• GPU-capable computer for training and testing the CNN.

Software & Tools

• ImageJ: Measures ring width, spot area, and intensity features from dried-blood-spot images.
• Python: Organizes images, trains the model, and runs the analysis pipeline.
• TensorFlow: Builds and tests a CNN for hemoglobin prediction.
• scikit-learn: Splits data, scores classification results, and checks model performance.
• Google Colab: Runs training code in the cloud when your computer is not powerful enough.

Experiment Steps

1. Define the prediction target, such as hemoglobin as a number or anemia as a yes or no label.
2. Choose the image features and sample types you will compare, then decide what counts as one data point.
3. Plan your reference data source and the validation strategy so your model can be checked on unseen samples.
4. Design a consistent imaging setup so lighting, focus, and scale stay the same across samples.
5. Set up a comparison between a simple baseline model and the CNN so you can judge whether the neural network adds value.
6. Plan error analysis before you start, so you can inspect failure cases by salinity, hematocrit, or paper type.

Common Pitfalls

• Letting lighting change between images, which makes the ring intensity look like a biological effect when it is really a camera effect.
• Mixing samples with different paper types without tracking them, which can hide the real link between spot shape and hemoglobin.
• Training on too few spots, which lets the CNN memorize the images instead of learning the pattern.
• Using the same samples for training and testing, which gives fake accuracy.
• Ignoring salinity and hematocrit together, which can make the model blame the wrong variable for a change in the coffee-ring pattern.

What Makes This Competitive

A stronger version of this project would test more than one sample condition and prove the model still works when the image changes a bit. You could compare a CNN against simpler image features, then show where each one fails. You could also use a strict train, validation, and test split, plus confidence intervals or error bars, instead of only reporting accuracy. That kind of careful analysis makes the project feel like real diagnostic research, not just image classification.

Project Variations

• Use published dried-blood-spot images from different patient groups and test whether the model generalizes across datasets.
• Compare grayscale features, color features, and CNN outputs to see which approach best predicts hemoglobin.
• Change the paper substrate or drying environment and test how much the spot pattern shifts before the model breaks down.

Learn More

• PubMed: Search review articles on dried blood spot analysis, anemia screening, and hemoglobin estimation.
• NIH: Search for papers and background material on anemia, blood biomarkers, and diagnostic methods.
• NASA Open Data Portal: Find image analysis examples and free datasets for practicing machine learning workflows.
• MIT OpenCourseWare: Look for free courses on machine learning, probability, and computer vision.
• ImageJ Documentation: Read the official guides for measuring image intensity and spot geometry.
• Peer-reviewed journals such as Biosensors and Bioelectronics: Search for studies on low-cost diagnostics and image-based blood testing.