GridSift automatically classifies raw MRI series from any hospital export — T1, T2, FLAIR, DWI, DCE — and sorts them into labeled folders, ready for research.
What is GridSift
Built for medical imaging researchers
Hospital imaging exports arrive unlabeled. Naming conventions differ by institution, vendor, and country — there's no universal standard to rely on.
Before any analysis begins, someone has to manually sort through hundreds of DICOM files. On large datasets, that's days of work before research even starts.
GridSift's physics-based filter reads acquisition parameters — TE, TR, flip angle, inversion time — to classify series automatically, regardless of how they're named.
A local LLM handles what the physics filter can't resolve. Uncertain series are always flagged for review, never silently assigned the wrong label.
Everything runs on your machine. Here's what that means in practice.
Fully local
No data leaves your machine. No API keys, no internet connection, no cost per run. Built for sensitive clinical data.
Physics-first
MRI acquisition parameters are the primary signal — so it works across institutions regardless of how series are named.
Uncertainty flagging
Anything the pipeline isn't confident about gets flagged, not guessed. Every decision is logged and auditable.
Sorted output
Results land in labeled subfolders — T1, T2, FLAIR, DWI, DCE, flagged_for_review — ready to hand off to your pipeline.
CSV export
Each classification is logged with a confidence score and decision path. Fully reproducible from run to run.
Multi-site ready
Tested across institutions, vendors, and naming conventions. No retraining, no reconfiguring between sites.
Before vs. After
What using GridSift actually changes
Without GridSift
Sorting hundreds of unlabeled DICOM files by hand — often 2 to 4 days before any analysis can start
One-off scripts that break the moment a new institution uses different naming conventions
Misclassifications that silently corrupt training data, with no record of what went wrong
No audit trail — preprocessing steps are impossible to reproduce across researchers
Every team member produces a slightly different folder structure
With GridSift
Hundreds of files classified and sorted in minutes, not days
Works across any institution's export without configuration or retraining
Low-confidence series get flagged — you only review what actually needs a second look
Every decision logged with a confidence score and the path that produced it
Same input always produces the same output — fully reproducible from run to run
Install GridSift
Free for all platforms. No account required.
Install GridSift
Requires Python 3.9 or later. Runs on Windows, macOS, and Linux.
GridSift uses a local LLM (Llama 3.2 3B) to classify ambiguous series descriptions. Install the model file below and place it inside the gridsift folder before running. This is a one-time 2GB install.
Download the Llama 3.2 3B model file (2GB) from the Downloads page and place it inside the cloned gridsift/ folder. The app looks for the model file there on startup.
gridsift/
app.py
Llama-3.2-3B-Instruct-Q4_K_M.gguf ← place here
... other files
3
Open GridSift and select your folders
Run python app.py from the gridsift folder. Click Select Folder to choose your DICOM dump folder. Click Select Output to choose where sorted files will be saved. Optionally type a Study Name to create a named subfolder.
4
Run classification
Click Run Classification. GridSift scans all DICOM files, groups them by series, and classifies each one. A progress bar shows real-time status. Results appear in the table as each series is classified.
# GridSift processes your data:
Found 847 files in 23 series
Pancreas → CT (0.99) · physics
Brain STIR → FLAIR (0.90) · physics
CBU_DTI_64D → T2 (0.85) · physics
RESTING_STATE → unknown (0.0) · flagged
5
Review flagged series and export
Series highlighted in red were uncertain — review those manually. Click Export Results as CSV to save a full log. Your output folder now contains sorted subfolders ready for your research pipeline.