---
title: Image Quantization and CT Windowing Explorer
emoji: 🩻
colorFrom: blue
colorTo: gray
sdk: docker
app_port: 8501
pinned: false
---

# Image Quantization and CT Windowing Explorer

An interactive Streamlit app for exploring how quantization and CT-style windowing change grayscale medical image display.

## Screenshots / GIF

### Main app overview

![Main app overview](screenshots/demo_overview.gif)


## Project Summary

This app compares two related but different operations on grayscale medical images:

1. Quantization reduces the number of stored gray levels and can introduce banding or reconstruction error.
2. Windowing changes only how a chosen intensity range is displayed and clips values outside that range.

The interface lets users adjust bit depth and window bounds, then compare four image states, a histogram, an error map, and summary metrics side by side.

## Local Run

```bash
cd medical-quantization-windowing-explorer
make setup
source .venv/bin/activate
make run
```

Or run without activating:

```bash
uv run streamlit run app.py
```

## Hugging Face Space URL

For the most stable viewing experience, use the direct app URL:
https://huggingkatze-ct-windowing-quantization-demo.hf.space

The Hugging Face project page is:
https://huggingface.co/spaces/HuggingKatze/ct-windowing-quantization-demo

## Repository Structure

```text
app.py
processing.py
metrics.py
utils.py
sample_images/
sample_images/builtin_samples/
docs/
requirements.txt
Makefile
Dockerfile
```

## Features

- Built-in CT-oriented examples from CT-RATE, LDCT-and-Projection-data, and RSNA PE
- Upload support for `PNG`, `JPG`, `JPEG`, and single-slice `DICOM`
- Automatic handling of grayscale images, including 3-channel grayscale copies
- Adjustable quantization bit depth and CT-style window bounds
- Slice selection for volumetric built-in samples
- Rotation control for uploaded DICOM slices
- Four synchronized views: original, windowed, quantized, and quantized + windowed
- Histogram, squared-error map, and summary metrics (`MSE`, `PSNR`, gray levels, entropy)
- Clickable comparison arrows synchronized with the metrics selector
- In-app explanatory notes for each main control

## Known Limitations

- The app is educational and not intended for clinical interpretation.
- Uploaded PNG/JPG images may already be restricted to 8-bit precision.
- The current version works on one 2D slice at a time.
- Uploaded files are limited to PNG/JPG and single-slice DICOM in the current app; NIfTI upload is not enabled in the active UI.
- The CT-RATE built-in sample is rotated 90 degrees clockwise in the app to match the expected viewing orientation.
- Public deployment should still respect the original dataset licenses and redistribution terms for bundled sample content.

## Interface Guide

- **Input source** switches between prepared built-in examples and your own uploaded data.
- **Built-in sample** lets you compare different CT-oriented datasets quickly.
- **Slice index** appears for volumetric built-in inputs and chooses the active axial slice.
- **Rotate uploaded image** appears for uploaded DICOM data when the slice orientation needs correction.
- **Quantization bit depth** controls how many gray levels remain after uniform quantization.
- The app initializes the quantization display at 4 bits by default, unless the loaded image has a lower maximum bit depth.
- **CT window preset** applies common lower/upper bound pairs such as lung or mediastinal ranges.
- **Window lower / upper bounds** and the numeric bound inputs define the exact displayed intensity interval.
- **Histogram** uses the selected window bounds as reference lines and is meant to show how quantization changes the intensity distribution, rather than to faithfully plot the post-windowed display mapping.
- **Visualization arrows** can be clicked to choose which pair of panels should be compared.
- **Metrics → Compare** is synchronized with the arrow graph and updates all downstream plots.

## Design Notes

See [docs/design_choices.md](docs/design_choices.md).