Smartphone Acne Fluorescence Imaging Project

ISEF Category: Biomedical Engineering

Subcategory: Biomedical Sensors and Imaging  ·  Difficulty: Advanced  ·  Setup: University Lab  ·  Time: Full Year

The Hook

Your skin can glow back at you, if you shine the right light on it. Some acne bacteria make porphyrins, pigments that fluoresce under UV light. That gives you a way to turn an invisible skin signal into measurable image data. If you pair that signal with machine learning, you can test whether a cheap phone rig can help grade acne severity.

What Is It?

This project studies a simple idea with a lot of tech packed into it. When you shine UV light on skin, some compounds emit visible red or orange light. That glow comes from porphyrins, which are molecules linked to acne bacteria such as Cutibacterium acnes (the newer name for P. acnes). Think of it like using a blacklight to find hidden ink, except the ink comes from biology.

The imaging part matters just as much as the biology. A smartphone camera can record the glow, and an IR-pass filter can change what the camera sees by blocking visible light and passing infrared. A multispectral setup means you capture more than one kind of image or channel, then compare them. A vision transformer is a machine learning model that reads image patterns and can be trained to predict acne severity from labeled photos, including public datasets such as Acne04.

The core challenge is not just seeing a bright spot. You need to decide whether fluorescence signal, skin tone, lighting, camera settings, and lesion count all affect the final score. That makes the project a mix of optics, image analysis, and data science.

Why This Is a Good Topic

This is a strong science fair topic because you can test a clear signal, build a cheap sensor, and compare your results against a labeled dataset. You are not guessing in the dark. You are measuring how well fluorescence imaging tracks acne features and whether a phone-based system can sort skin images in a useful way. That gives you a real engineering problem and a real analysis problem. You can also learn imaging calibration, data labeling, model evaluation, and bias checks, which are all useful research skills.

Research Questions

• How does UV wavelength affect the fluorescence signal from acne-prone skin regions?
• What is the effect of an IR-pass filter on the contrast between acne lesions and nearby skin?
• Does a smartphone-based fluorescence setup predict acne severity labels better than visible-light photos alone?
• To what extent do skin tone and lighting conditions change the measured fluorescence intensity?
• Which image features, lesion count, brightness, color ratios, or texture, best match Acne04 severity labels?
• How does a vision transformer compare with a simpler CNN for acne severity classification on the same image set?

Basic Materials

• Smartphone camera with manual exposure control.
• 365 nm or similar UV-LED light source.
• IR-pass filter that fits or can be held over the camera.
• Tripod or phone mount for fixed framing.
• Dark backdrop or blackout box to reduce stray light.
• White balance card or neutral gray card for calibration.
• Measuring tape or fixed-distance spacer.
• Consent form and image storage plan for any human-subject photos.
• Notebook or spreadsheet for image labels and observations.

Advanced Materials

• Spectrometer or calibrated light meter for checking LED output.
• Multispectral filter set or narrowband optical filters.
• Cross-polarization filters for reducing glare.
• Color calibration target for imaging correction.
• Computer with GPU access for model training.
• Python environment with image analysis libraries.
• Annotation software for lesion labeling.
• Secure storage for de-identified human skin images.
• Optional dermatoscope or clinical imaging attachment for comparison.

Software & Tools

• Python: Handles image preprocessing, feature extraction, and model training.
• ImageJ: Measures brightness, color channels, and region-based signal in skin images.
• Jupyter Notebook: Helps you document analysis steps and visualize results as you work.
• TensorFlow: Trains and evaluates vision transformer or CNN models on labeled images.
• scikit-learn: Splits data, scores classification results, and checks model performance.

Experiment Steps

1. Define the exact signal you will measure, such as fluorescence intensity, lesion count, or model accuracy.
2. Choose one imaging setup and lock down camera distance, lighting geometry, and filter placement.
3. Plan a calibration method that lets you compare images across days and across skin types.
4. Build a labeling system for acne severity that matches your dataset and stays consistent.
5. Decide how you will compare a physics-based signal analysis against a machine learning model.
6. Set evaluation metrics before you collect or train, so your results answer one clear question.

Common Pitfalls

• Using automatic phone exposure, which changes brightness between images and ruins comparisons.
• Letting ambient room light leak into the setup, which weakens fluorescence contrast.
• Mixing up visible glow with skin reflection, which makes porphyrin signal look stronger than it is.
• Training on unbalanced acne labels, which causes the model to favor the most common severity class.
• Ignoring skin tone and camera processing differences, which can hide bias in the final score.

What Makes This Competitive

A stronger version of this project does more than run a model on photos. It tests whether the fluorescence signal adds value beyond plain RGB imaging, and it checks that claim with clean controls. You can also compare multiple model types, measure performance across skin tones, and report false positives, false negatives, and class imbalance effects. That kind of analysis turns a simple imaging demo into a real engineering study.

Project Variations

• Test whether visible-light acne photos or fluorescence photos give better severity predictions on the same dataset.
• Compare UV-based imaging on oily skin versus dry skin to see whether surface properties change the signal.
• Add a color calibration target and see whether calibration improves model accuracy and repeatability.

Learn More

• PubMed: Search review articles on acne imaging, porphyrins, and optical diagnostics.
• NIH: Find background on acne biology and skin imaging in public health and research pages.
• NASA Open Science Data Repository: Look for image analysis examples and machine learning workflows you can adapt.
• ImageJ Documentation: Learn how to measure intensity, color channels, and regions of interest.
• MIT OpenCourseWare: Search for free materials on image processing, computer vision, and machine learning.