Spaces:
Sleeping
title: BioMedVisionLab
emoji: 🧬
colorFrom: blue
colorTo: green
sdk: docker
app_port: 7860
pinned: false
license: mit
BioMedVisionLab
A deployable biomedical visual evaluation workbench for image retrieval, super-resolution evaluation, and contact-map visualization.
BioMedVisionLab is a research prototype for inspecting biomedical image-like data through an interactive dashboard. It supports chest X-ray retrieval benchmarking, super-resolution baseline/model-output comparison, and Hi-C/contact-map visualization.
The goal is not diagnosis. The goal is visual evaluation: helping researchers inspect model behavior, compare enhancement outputs, identify failure cases, and export experiment summaries.
Live Demo
Try the deployed app here:
https://huggingface.co/spaces/Mahrufa/BioMedVisionLab
Hosted demo mode is intended for small uploaded samples. Large datasets, full benchmarks, and GPU-heavy workflows should be tested locally or on suitable research infrastructure.
Why This Project Matters
Biomedical AI models often produce embeddings, enhanced images, or reconstructed matrices, but researchers still need clear ways to inspect whether those outputs are meaningful.
BioMedVisionLab helps with:
- Inspecting retrieval behavior instead of only looking at final numbers.
- Comparing baseline vs model-enhanced biomedical images with visual and quantitative metrics.
- Visualizing image-like biological matrices, such as Hi-C/contact maps, for resolution-style analysis.
BioMedVisionLab is designed as a reusable visual evaluation workbench for biomedical image-like data, including radiology images, microscopy/chemical imaging outputs, pathology-style images, and genomic contact maps.
Core Modules
| Module | What it does | Research value |
|---|---|---|
| CXR Retrieval | Retrieves visually similar chest X-rays using encoder embeddings. | Compares generic vs CXR-specific visual representations. |
| Super-Resolution Demo | Simulates low-resolution images, reconstructs them with bicubic interpolation, and compares metrics. | Provides a transparent baseline for biomedical image enhancement workflows. |
| Grant Alignment Lab | Visualizes synthetic or uploaded contact maps and compares low-resolution vs upscaled matrices. | Bridges the dashboard toward Hi-C/chromatin contact-map analysis. |
Key Features
- Upload-first demo mode for biomedical images and contact maps.
- Chest X-ray retrieval using encoder embeddings.
- Encoder comparison support for:
- ResNet18 ImageNet baseline
- ResNet50 ImageNet baseline
- TorchXRayVision DenseNet121 CXR encoder
- Retrieval metrics including Precision@K, Top-1 match, label distribution, and mean cosine similarity.
- Out-of-domain warning for non-CXR uploads.
- Super-resolution baseline evaluation using PSNR, SSIM, and MSE.
- Optional upload of model-enhanced images for comparison against bicubic baseline.
- Synthetic and uploaded contact-map visualization using CSV/NPY inputs.
- Hi-C/contact-map metrics including matrix PSNR, SSIM, MSE, and difference maps.
- CSV exports for experiment tracking and method comparison.
- Docker-based deployment on Hugging Face Spaces.
Screenshots
Add screenshots or GIFs here after deployment.
Quick Start
Clone the repository:
git clone https://github.com/mahrufa-binta-ali/BioMedVisionLab.git
cd BioMedVisionLab
Create and activate a virtual environment.
On Windows:
python -m venv venv
venv\Scripts\activate
On macOS/Linux:
python -m venv venv
source venv/bin/activate
Install dependencies:
pip install -r requirements.txt
Run the app:
python -m streamlit run app.py
Optional: Build a Local CXR Index
The deployed demo can run in upload-first mode. For local CXR retrieval benchmarking, place a chest X-ray dataset under:
data/chest_xray/
train/
NORMAL/
PNEUMONIA/
test/
NORMAL/
PNEUMONIA/
Then build embeddings for one or more encoders:
python build_index.py --encoder resnet18_imagenet
python build_index.py --encoder resnet50_imagenet
python build_index.py --encoder torchxrayvision_densenet121
Embeddings are saved under:
embeddings/
resnet18_imagenet/
resnet50_imagenet/
torchxrayvision_densenet121/
Large datasets and embedding files are intentionally excluded from Git.
Deployment
This project is deployed using Docker on Hugging Face Spaces.
The Space uses:
sdk: docker
app_port: 7860
The Dockerfile runs the Streamlit app on port 7860.
To deploy manually:
- Create a new Hugging Face Space.
- Select Docker as the SDK.
- Push this repository to the Space.
- Hugging Face will build the Docker image and launch the app.
Example push command:
git remote add space https://huggingface.co/spaces/Mahrufa/BioMedVisionLab
git push space main
If the Space already has an initial commit and rejects the push, use:
git push space main --force
CSV Reports
BioMedVisionLab supports CSV exports for experiment tracking.
Single reports describe the current selected example. Batch benchmark reports export many rows and are intended for method comparison.
| Report | Purpose |
|---|---|
| Single-query retrieval report | Inspect one query and its retrieved cases. |
| Batch retrieval benchmark | Compare retrieval quality over many query images. |
| Batch retrieval summary | Summarize Precision@K, Top-1 match, and class-level retrieval behavior. |
| Single-image SR report | Record PSNR, SSIM, and MSE for one image. |
| Batch SR benchmark | Evaluate reconstruction metrics over many images. |
| Hi-C matrix evaluation report | Record contact-map enhancement metrics for a matrix experiment. |
Intended Use
BioMedVisionLab is intended for:
- Research prototyping
- Visual model evaluation
- Biomedical image-like data inspection
- Retrieval benchmarking
- Super-resolution baseline comparison
- Contact-map visualization experiments
- Teaching and portfolio demonstration
It is not intended for clinical diagnosis, patient-facing use, or medical decision-making.
Limitations
- This is a research prototype, not a diagnostic system.
- Retrieval labels come from nearest dataset examples and should not be interpreted as predictions.
- Retrieval confidence is not clinical confidence.
- Bicubic super-resolution is a transparent baseline, not a trained AI enhancement model.
- Synthetic Hi-C/contact maps are interface demonstrations, not biological results.
- Uploaded files in hosted demo mode are processed temporarily during the session.
- Large biomedical datasets may require local execution or GPU-backed infrastructure.
- The current hosted demo is designed for small samples, not full-scale clinical or genomics pipelines.
Research Roadmap
Planned extensions include:
- CXR-specific foundation-model encoders.
- Grad-CAM or model-based visual grounding.
- Report-aware CXR retrieval using radiology text.
- Microscopy/pathology super-resolution benchmarks.
- Diffusion-based biomedical image enhancement comparison.
- Real
.cool/.mcoolHi-C contact-map support. - Batch evaluation dashboards for multiple models and datasets.
Tech Stack
- Python
- Streamlit
- PyTorch
- torchvision
- TorchXRayVision
- NumPy
- pandas
- scikit-image
- Altair
- Docker
- Hugging Face Spaces
Project Structure
BioMedVisionLab/
│
├── app.py
├── build_index.py
├── requirements.txt
├── Dockerfile
├── README.md
├── .gitignore
│
├── .streamlit/
│ └── config.toml
│
├── docs/
│ ├── TECHNICAL_NOTES.md
│ └── ROADMAP.md
│
├── data/ # ignored by Git
├── embeddings/ # ignored by Git
└── assets/ # optional screenshots/GIFs
Disclaimer
BioMedVisionLab is a research visualization prototype. It is not approved for clinical use and should not be used for diagnosis, treatment planning, or medical decision-making.
Uploaded images are not classified by the app. Labels shown in retrieval modules come from nearest reference examples and are only used for visualization and evaluation.
Author
Mahrufa Binta Ali
GitHub: mahrufa-binta-ali
Project: BioMedVisionLab