Potestates/DeepSparse-25v

DeepSparse — 25-view finetuned checkpoints

Finetuned checkpoints of DeepSparse for 25-view (180°) sparse-view CT reconstruction on four anatomies. All models start from the official pretrained weights (pretrain/ep_700.pth from HajihajihaJimmy/DeepSparse) and use the official two-stage finetuning protocol.

The folder layout matches the official repo, so the checkpoints drop into the codebase's logs/ directory.

Results (official test splits, 256³, 3D PSNR / SSIM)

Checkpoint	Train cases	Test cases	PSNR (dB)	SSIM (×10⁻²)
pelvis+25v+n250+s2 (PENGWIN)	60 (all)	30	30.96 ± 2.63	90.31 ± 3.20
luna+25v+n250+s2 (LUNA16)	250 / 738	100	32.47 ± 1.05	92.33 ± 1.83
abdomen+25v+n250+s2 (PANORAMA)	250 / 1244	600	29.79 ± 2.11	89.52 ± 3.21
tooth+25v+n250+s2 (ToothFairy)	250 / 343	75	34.03 ± 1.15	94.29 ± 1.53

Mean ± std over test cases, from the official code/evaluate.py. Per-case results are in each folder's results_1.0x.csv.

For reference, paper Table III (10 views, full training sets): pelvis 29.03 / 90.27, LUNA16 31.86 / 91.41, abdomen 29.42 / 88.36, tooth 31.79 / 92.50. On the full LUNA16 test set, the official 10-view checkpoint gives 31.79 / 91.46 on our preprocessed data.

Setup

• Code: xmed-lab/DeepSparse @ a055aba3bcb5732a68f89cc0bf3ee6fbfbb1b1e4
• Views: 25 input views, uniformly over 180°, taken from a 300-view projection cache
• Stage 1: --num_views 30 --vq_w 0.1, resumed from pretrain/ep_700.pth, 400 epochs
• Stage 2: --num_views 30 --min_views 25 --random_views --vq_w 1.0 --safely_load --freeze_ft, resumed from stage-1 ep_400.pth, 400 epochs
• Batch size 2, lr 1e-4, weight decay 1e-3, no LR schedule, 1 GPU per job
• Environment: PyTorch 2.7.1 + CUDA 12.8, TIGRE 3.1.3

Differences from the paper protocol

1. Training set size: at most 250 training cases per dataset (the paper uses the full training sets). Test splits are the official ones. The exact case lists are in splits/<DATASET>/meta_info.json.
2. Dense views: num_views=30 during training, so stage 2 has a teacher/student view ratio of 1.2× (official: 24 dense views vs 6/8/10).
3. Environment: newer PyTorch and TIGRE than the official (PyTorch 1.13, TIGRE 2.3), needed for Blackwell GPUs.
4. ToothFairy config: the official meta_info.json references an unreleased config+new.yaml / processed+new/. We use the repo's public config.yaml. The official tooth 10-view checkpoint reproduces the paper numbers with this config (31.89 / 92.74).

Files

<dataset>+25v+n250+s2/
  ep_400.pth          final stage-2 checkpoint (use this)
  config.yaml         config saved by train.py (note: min_views stays 10 here)
  results_1.0x.csv    per-case test PSNR/SSIM + average
  train.log           stage-2 training log (args + val curve)
  train_s1.log        stage-1 training log
configs/
  finetune_s1_n250.yaml, finetune_s2_n250.yaml   training configs (root_dir ./data_n250)
  eval_25v_n250.yaml                             evaluation config (min_views: 25)
splits/<DATASET>/meta_info.json                  train/eval/test case lists used

Usage

# inside a DeepSparse checkout with data prepared as in the official README
hf download Potestates/DeepSparse-25v --local-dir ./logs
cp logs/configs/eval_25v_n250.yaml configs/generated/   # set root_dir to your data root

python code/evaluate.py --name luna+25v+n250+s2 --epoch 400 --dst_name luna \
  --split test --num_views 25 --cfg_path configs/generated/eval_25v_n250.yaml \
  --out_res_scale 1.0

Evaluation must use min_views: 25. The config.yaml saved inside each checkpoint folder keeps min_views: 10, so use configs/eval_25v_n250.yaml instead.

License

The DeepSparse code is MIT. These weights are derived from datasets with their own terms. Check each dataset's license before use:

• PANORAMA (abdomen): CC BY-NC 4.0 (non-commercial)
• PENGWIN (pelvis): CC BY 4.0
• LUNA16 (lung), ToothFairy (tooth): see the original dataset terms

Please cite the DeepSparse paper and the datasets if you use these checkpoints.