DeltaScan

Inspiration

Judge spoke to us about how specificity can vastly improve performance with compression and efficiency, as information relevancy becomes more concise as generalization decreases. CT/MRI scans tend to be difficult to book and expensive, given the amount of time that goes into drawing conclusions from the sheer amount of data.

What it does

DeltaScan uses a novel compression algorithm to find the "delta" or the differences across CT images which may have hundreds to thousands of images, costing significant tokens to process for LLM input. Our algorithm results in a delta image which saves up to 80-95% of tokens for LLM input. An LLM then analyzes this delta image and returns a diagnosis of the CT scan using a reference database that ensures accuracy.

How we built it

We used numpy for the backend algorithm, FastAPI to make into a public REST API, and NextJS for the frontend. We used Gemini AI to analyze the results.

Challenges we ran into

The biggest challenge was preventing error accumulation in our delta encoding algorithm - we were accidentally compounding compression artifacts across 200+ slices, which we fixed by comparing against original data instead of reconstructed intermediates We also ran into challenges with things like the DICOM file structure, git push failures due to HTTP 408 timeouts, processing speeds (went from minutes to seconds per scan), and quality v.s. compression tradeoff

What we learned

After working through the project, we learned about the different types of compression algorithms, specifically for images. Going through each unique model and seeing if it suited our needs was eye-opening and was fun to learn.

Built With

• fastapi
• nextjs
• python
• typescript