Brain lesion maps · UCSF

Simulate

DICOM folder …/pilot/input Found 6 patients · 7 cases This computer ready for all site steps ✓

Next: 1 Find cases

Waiting for you

Finds each MRI and its tumor contours in your DICOM folder, including inside zip files.

Progress0 %

Done0 of 6 patients

Time left≈ 1 min

Speed–

Run the next steps automatically. The Console stops only when it needs a decision from you.

Your pipeline

✓ checked● not yet⟳ needs re-run👤 needs your decision

Patients

Pseudonyms only. Names stay in your private clinical workbook.

This computer

Part	This computer	Status
Processor	Workstation CPU · 32 cores (Docker may use 24)	✓
Memory	128 GB (Docker may use 96 GB)	✓
GPUs	2 × NVIDIA GPU · 48 GB each · visible from containers	✓
Disk (workspace)	4 TB free	✓
Docker	Docker Engine 28 · Linux containers · NVIDIA runtime	✓
Performance mode	Balanced: 24 CPUs, 2 GPUs (one case per GPU at a time)	change

Step	Needs	Here	Estimated time for 2,000 patients
1 Find cases	2 CPUs · 2 GB · 10 GB disk	✓ ready	≈ 25 min
2 Name structures	1 CPU · 1 GB	✓ ready	≈ 2 min + your review
3 De-identify	4 CPUs · 4 GB · disk ≈ size of the MRIs	✓ ready	≈ 40 min
4 Map to MNI	NVIDIA GPU ≥ 8 GB (required) · 32 GB RAM · ≈ 0.5 GB per patient	✓ ready (2 GPUs)	≈ 2 days 12 h
5 Check quality	4 CPUs · 8 GB	✓ ready	≈ 30 min + your review
6 Add clinical data	Excel or Numbers to fill the workbook	✓ ready	your time
7 Prepare submission	1 CPU · disk ≈ 5 MB per patient	✓ ready	≈ 3 min

Step 4 runs only on an NVIDIA GPU, so results match the UCSF atlas and stay fast. On a computer without one, step 4 shows ✗ and offers to send the de-identified release to UCSF for processing.

Containers

Each part of the pipeline is a Docker container shipped inside the Console and checked against its SHA-256 before use. Containers never have network access, and your DICOM folder is always mounted read-only.

How it fits together

The goal is a probabilistic map of where brain metastases occur, in the shared MNI brain space. Your computer turns your DICOM data into one lesion mask per tumor in MNI space, plus a clinical table with no identifiers. UCSF combines the submissions of all sites into the maps.

Orange area: identifiable data. It never leaves this computer. Green area: de-identified data, checked after every step. Purple area: the only thing that leaves your institution, shown to you file by file before it is created.

Between two steps, the Console checks the data with an independent checker. A step can only start when the data it needs passed that check and this computer has what the step needs.

You are asked for a decision four times: confirm structure names (2), confirm cases and approve the release (3), confirm quality flags (5) and fill the clinical workbook (6). Everything else runs on its own.