Plant-Based Tissue Scaffolds with CNN Pore Analysis

ISEF Category: Biomedical Engineering

Subcategory: Biomaterials and Regenerative Medicine  ·  Difficulty: Intermediate  ·  Setup: Home Setup  ·  Time: 1 to 2 Months

The Hook

Researchers have grown human cells on decellularized spinach leaves where the leaf veins act like blood vessels. You can run the same trick with banana peels, onion skins, and eggshell membranes from your kitchen. Two polarizer films from old 3D-movie glasses turn your phone into a microscope that sees protein fiber alignment. A small neural network then learns to predict pore sizes from those images.

What Is It?

Decellularization means washing away all the living cells from a tissue and keeping the structural scaffold behind. The scaffold is the fiber mesh that gave the tissue its shape. Plants and eggshell membranes have meshes that look surprisingly like animal extracellular matrix under a microscope.

You can mimic the standard surfactant wash using dish soap (sodium dodecyl sulfate is the active ingredient), dilute bleach, and ethanol. After several rinses the tissue turns pale and translucent because the cells are gone. What remains is mostly cellulose or collagen-like fibers.

Polarized light passes through aligned fibers differently than through random ones. By stacking two polarizer films and rotating one, you can produce a contrast map of fiber alignment on your phone camera. A convolutional neural network (CNN) is a vision model that learns image-to-number mappings, here from optical image to pore-size distribution.

Why This Is a Good Topic

Decellularized scaffolds are a real-world topic in tissue engineering, and the food-grade version is reproducible without lab access. Pore size controls whether cells can move into a scaffold, so the prediction target is biologically meaningful. You will learn surfactant chemistry, polarized-light microscopy, and how to train a small CNN with limited data.

Research Questions

• How does wash time change the residual cell-debris signal in each tissue?
• What is the effect of bleach concentration on fiber-alignment retention?
• Does polarized-light contrast predict pore size measured directly with ImageJ?
• To what extent does a CNN generalize across banana, onion, and eggshell scaffolds?
• Which tissue source gives the most reproducible pore distribution?
• How does drying method affect measured pore size?
• What is the effect of training-set size on CNN regression error?

Basic Materials

• Banana peels, onion skins, and chicken eggshells.
• Unscented dish soap (SDS surrogate).
• Household bleach diluted to label specifications.
• Ethanol (drugstore 70 or 91 percent).
• Two linear-polarizer film sheets.
• Smartphone with macro clip-on lens.
• Petri dishes or plastic culture trays.
• Digital kitchen scale.

Advanced Materials

• Lab-grade SDS solution.
• Inverted phase-contrast microscope.
• DAPI stain to confirm residual DNA removal.
• Lyophilizer for porosity quantification.
• GPU for CNN training.

Software & Tools

• ImageJ: Manually measures pore sizes to build ground-truth labels.
• TensorFlow or PyTorch: Trains the CNN regression model.
• Python (scikit-learn): Splits data and computes regression metrics.
• OpenCV: Preprocesses smartphone images and normalizes polarization contrast.

Experiment Steps

1. Choose the tissue sources you will compare and lock a single decellularization protocol per tissue.
2. Build a polarized-light imaging jig with a fixed light source so contrast is comparable between sessions.
3. Label a starter set of images with hand-measured pore sizes in ImageJ before touching the model.
4. Decide your train-validation-test split so the CNN never sees test images during training.
5. Plan controls that rule out tissue-color artifacts driving the CNN prediction.
6. Compare CNN predictions against ImageJ ground truth on a held-out tissue type.

Common Pitfalls

• Over-bleaching the scaffold, which destroys the fibers you wanted to image.
• Photographing under changing room light, which makes polarization contrast drift between sessions.
• Training a CNN on fewer than 50 labeled images per tissue, which hides overfitting.
• Mistaking residual cell debris for fiber structure in polarized images.
• Skipping a hand-labeled validation set and reporting only training loss.

What Makes This Competitive

Move from a class-level demo to a competitive project by adding a cross-tissue generalization test, a comparison against a published decellularization quality metric like residual DNA content, and an explainability layer such as Grad-CAM on the CNN. Reporting confidence intervals on pore-size predictions and a calibration plot against ImageJ measurements adds rigor.

Project Variations

• Swap eggshell membrane for fish scales descaled from grocery-store fish and compare collagen alignment.
• Replace the CNN with a vision transformer and compare sample efficiency on small datasets.
• Use the scaffold as a substrate for yeast adhesion and link pore size to colony density.

Learn More

• PubMed: Search decellularized plant scaffolds and Gershlak spinach leaf paper.
• NIH PubMed Central: Open-access reviews on extracellular matrix scaffolds.
• TensorFlow tutorials: Official image regression walkthroughs.
• ImageJ User Guide: Built-in documentation on particle analysis.
• MIT OpenCourseWare: Course 20.109 Laboratory Fundamentals in Biological Engineering.