spine-mri-vision-model-training

Spine MRI vision model training weights

Checkpoints for the computer-vision-model-training package in ClinicalDCM2BIDS (Vision-Model-Training branch).

Two training rounds share this repo. Round 1 sits at the top level; round 2 is under round2/. The two use different protocols — read the round 2 notes below before comparing numbers across them.

Round 1 — 13 supervised 2D backbones

Layout matches models/full_all13/ (52 final.pt files, plus fold0..4.pt):

• 13 backbones: alexnet, googlenet, resnet18, densenet121, efficientnet_b0, convnext_tiny, efficientnet_v2_s, mobilenet_v3_small, vgg11, coatnet_0, maxvit_tiny, swin_tiny, brain_seq_linear_probe
• Per model: separate/{type,voi,acq}/final.pt and multihead/final.pt
• Protocol: 5-fold cross-validation plus a final train/val fit

Round 2 — 3D CNNs and frozen 2D foundation probes

24 final.pt files under round2/, adding two model families the round 1 roster did not cover.

Track A — 3D CNNs, full-volume input [1, 96, 128, 128]:

Track B — frozen mid-slice foundation encoders, input [3, 224, 224], encoder frozen with only a linear head trained:

Per model: separate/{type,voi,acq}/final.pt and multihead/final.pt.

Round 2 protocol differences

Three things differ from round 1 and matter when comparing results:

1. No cross-validation. Round 2 runs a single 80/10/10 train/val/test fit per model and task, so there are no fold*.pt files — only final.pt. Round 1 numbers are 5-fold averages; round 2 numbers are single-split.
2. densenet121_3d trains from scratch. MONAI raises NotImplementedError for 3D pretrained DenseNet weights, so it falls back to random init. Comparing it against the two MedicalNet-pretrained ResNets is transfer learning versus from-scratch, not an architecture comparison.
3. Track B checkpoints embed the frozen encoder. Each file contains the full model, so most of its size is a copy of public pretrained weights; only the linear head is novel.

Checkpoint dicts hold {"model": state_dict, "epoch": int, "val": metrics}.

Download

# Round 1
hf download klasinlapaprechar/spine-mri-vision-model-training \
  --local-dir models/full_all13 --repo-type model

# Round 2 only
hf download klasinlapaprechar/spine-mri-vision-model-training \
  --include "round2/*" --local-dir models/round2 --repo-type model

Code / metrics

Training code and results CSVs/figures are on GitHub. This Hub repo is weights only — it contains no patient data, no per-scan predictions, and no identifiers.