Step 1 · Data
What are you working on today, Dr. Chen?
Drop a labeled dataset and we'll handle the rest. Pick a task, start training, and get a validated model back.
Drop DICOM / NIfTI / CSV folder here
or click anywhere to load the sample chest CT set
chest-ct-100cases.zip
0%
or start from a sample:
chest-ct-100cases.zip
De-identified · BAA covered
Label distribution
Step 2 · Task
Pick the task
Choose what best matches your clinical question. You can change this later.
Selected:
Step 3 · Configure
Smart configuration
Clique fingerprinted your dataset, read the cluster's hardware inventory, and pre-filled every setting for the GPUs you actually have. Here's what it chose.
Why this configuration? Auditable decision tree · nnU-Net-inspired
Every choice below is driven by two things the platform already measured for you: the dataset fingerprint and the available hardware. Each answer is computed from those inputs, logged, and reproducible.
-
1 Is the dataset 3D or 2D?3D: 100 CT volumes, median shape 512×512×220.→ Picks a volumetric architecture family (UNet3D) over 2D slice-wise baselines.
-
2 Is the dataset large enough for the full nnU-Net?No: 100 cases is below the 200-case threshold (heuristic H-3).→ Downgrades to the UNet3D-small variant to avoid over-parameterization.
-
3 What batch size fits in available VRAM?Batch = 2: solved from 24 GB A30 VRAM minus ~3.1 GB headroom at median shape.→ Keeps gradient accumulation off the critical path; no OOM at epoch 0.
-
4 Is mixed-precision safe for this task?Yes: segmentation on CT + A30 = ~2.1× speedup, <0.1% Dice impact in published benchmarks.→ Enables Mixed FP16 by default; falls back to FP32 automatically if NaN losses are detected.
-
5 Which augmentation policy?Elastic + intensity jitter: small 3D CT dataset, single modality.→ Matches nnU-Net's default 3D CT policy. Skips color / mixup (not meaningful on CT).
-
6 How to validate?5-fold CV: N < 200 and we want a variance estimate for the handoff report.→ Single-split would be misleading at this sample size.
Every branch of this tree is recorded in the audit log with the dataset hash, hardware ID, and resolved value, so IT and compliance can reproduce the run on demand.
Architecture
UNet3D-small
Lightweight 3D variant optimized for a 100-case training set.
Batch size
2
Fits comfortably in 24GB GPU with 3.1GB headroom.
Precision
Mixed FP16
2.1× speedup vs FP32, <0.1% accuracy impact.
Augmentation
Elastic + intensity jitter
Standard for 3D CT with small training sets.
Validation split
5-fold cross-validation
Recommended for datasets < 200 cases.
Estimated runtime
4 h 12 min
On 1× NVIDIA A30 (detected in your cluster).
Advanced options Researchers only. Clinicians can safely ignore this.
We'll provision a container and stream progress below.
Step 4 · Train
Watch training live
Nothing is running yet. Click Start training above and the lifecycle, chart, and log below fill in live.
Container provisioned in 4.8s. It shuts down automatically when training completes, so the GPU goes back to the cluster instead of idling.
-
ConfiguringContainer provisioned · image: clique-runner:v4.2
-
TrainingEpoch 0 / 250 · no metrics yet
-
Validating5-fold cross-validation
-
ExportingONNX + versioned artifact
-
ReleasedAvailable in Model Registry
Loss & validation dice
Train loss
Val dice
GPU util
n/a
Power
n/a
Elapsed
00:00
ETA
n/a
Live log
clique-runner:v4.2
No job yet. The live log streams here once training starts.
Model ready: pneumonia-seg-v1
Validation complete. Artifact pushed to your Model Registry.
Step 5 · Deploy
Your model is ready
Review the artifact, hand it to IT, or deploy it yourself, with every step audited.
pneumonia-seg-v1
Sensitivity
0.932
Specificity
0.908
AUC
0.961
Val Dice
0.891
Artifact
148 MB · ONNX
SHA-256
a7f4c2…e9b1
Lab book
Recent experiments
Every experiment is a container in some stage of its lifecycle: configured, trained, validated, exported, released, or archived.
| Name | Task | State | Dataset | GPU-hours | Accuracy | Updated |
|---|