AI Low-Dose Pediatric CT Super-Resolution

ISEF Category: Biomedical Engineering

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

The Hook

CT scans can save lives, but radiation adds risk, especially for kids. What if a computer could clean up a noisy scan so doctors could use far less dose? That is the core idea behind AI super-resolution for low-dose CT. You can test whether the image still keeps the details that matter.

What Is It?

This project asks whether an AI model can turn a noisy, low-dose CT image into something closer to a standard-dose image. Think of it like sharpening a blurry photo, but with medical scans where tiny details can affect diagnosis. The model learns patterns from examples, then predicts what the cleaner image should look like.

A diffusion model is one type of AI that learns how to remove noise in a step-by-step way. In this project, you would fine-tune that model on pediatric CT data and then test whether it preserves lesion conspicuity, which means how easy a lesion is to see. You are not trying to make a fake image. You are testing whether the model improves visibility without erasing or inventing medical detail.

Why This Is a Good Topic

This is a strong science fair topic because you can measure it in clear numbers. You can compare image quality, noise, contrast, structural similarity, and lesion visibility across different dose levels. The project connects to a real clinical problem, since pediatric imaging needs lower radiation whenever possible. You can also learn how AI image restoration, evaluation metrics, and medical imaging tradeoffs work together.

Research Questions

• How does simulated dose reduction affect lesion conspicuity in pediatric CT images after diffusion-model reconstruction?
• What is the effect of different noise injection levels during training on image quality metrics such as SSIM and PSNR?
• Does fine-tuning on pediatric CT data improve reconstruction quality more than using a generic pretrained model?
• To what extent does the model preserve edge detail in anatomically complex regions after low-dose degradation?
• Which evaluation metric, SSIM, PSNR, or a lesion visibility score, best matches human judgment of image usefulness?
• How does the model perform on scans from different body regions, such as head, chest, or abdomen?

Basic Materials

• Laptop or desktop computer with a modern GPU if possible.
• Python 3 with Jupyter Notebook.
• TCIA pediatric CT dataset access.
• Open-source deep learning framework such as PyTorch or TensorFlow.
• ImageJ or Fiji for quick visual checks.
• External storage for large imaging files.
• Spreadsheet software for organizing scan groups and results.
• Basic statistics tool or Python plotting library for analysis.

Advanced Materials

• Workstation with NVIDIA GPU and enough VRAM for diffusion training.
• Python environment with PyTorch, MONAI, NumPy, SciPy, and scikit-image.
• TCIA pediatric CT dataset access and data-use documentation.
• Medical image tools such as 3D Slicer or ITK-SNAP for inspection.
• ImageJ or Fiji for 2D slice analysis.
• A curated set of lesion or structure annotations, if available.
• Model checkpoint storage and experiment tracking logs.
• Cloud compute credits or institutional GPU access for larger runs.

Software & Tools

• Python: Runs the model training, data preprocessing, and evaluation code.
• PyTorch: Supports building and fine-tuning the diffusion model.
• MONAI: Provides medical imaging workflows and common loss functions.
• ImageJ: Helps you inspect slices and compare image quality by eye.
• 3D Slicer: Lets you view CT volumes and check whether details survive reconstruction.

Experiment Steps

1. Define the imaging question you want to test, such as whether lower-dose inputs still preserve visible lesions after reconstruction.
2. Choose one pediatric CT task and one evaluation target, then keep both fixed so your comparisons stay fair.
3. Build a clean data split so scans from the same patient do not leak across training and testing.
4. Decide how you will simulate low-dose noise and which quality metrics will score the outputs.
5. Plan a baseline model first, then compare your diffusion model against that baseline and against the original noisy input.
6. Set up human or rule-based review criteria so you can judge whether the model improves visibility without changing anatomy.

Common Pitfalls

• Mixing slices from the same patient across train and test sets, which makes the model look better than it really is.
• Using only PSNR or SSIM, which can miss cases where a lesion becomes harder to see.
• Training on too few scan types, which causes the model to fail on body regions it never learned.
• Injecting noise in a way that does not match real CT noise, which weakens the clinical meaning of your results.
• Skipping visual inspection of outputs, which lets hallucinated details or erased lesions hide inside average metric scores.

What Makes This Competitive

A competitive project would go beyond simple image cleanup. You could compare several dose levels, several reconstruction settings, and more than one evaluation metric. Strong entries also separate anatomy preservation from visual sharpness, so the model does not win by making images look smoother than they really are. If you can test whether the model helps some scan types more than others, your analysis becomes much stronger.

Project Variations

• Test the same pipeline on adult CT scans and compare whether pediatric data needs different fine-tuning.
• Swap the diffusion model for a U-Net or GAN baseline and compare image fidelity and lesion visibility.
• Analyze one organ system, such as lung or brain, and study whether the model keeps small structures intact.

Learn More

• The Cancer Imaging Archive (TCIA): Search for pediatric CT collections and read the dataset documentation on the TCIA site.
• PubMed: Search for review articles on low-dose CT reconstruction, diffusion models, and lesion conspicuity.
• NIH National Library of Medicine: Read background articles on CT imaging, radiation dose, and pediatric imaging safety.
• MONAI Documentation: Find open medical imaging tutorials and model examples on the MONAI project site.
• 3D Slicer Documentation: Learn how to inspect CT volumes and compare reconstructed images in 2D and 3D.