| tags: | |
| - ct-reconstruction | |
| - sparse-view-ct | |
| - x-ray | |
| - pytorch | |
| # NSP V3 strong-fusion Phase-1 prior | |
| This release contains the best Phase-1 checkpoint from the V3 strong X-ray | |
| fusion experiment. It was warm-started from the uploaded v27 prior and trained | |
| on 500 LUNA16 cases (450 train / 50 held out). | |
| ## Checkpoint | |
| - `best.pth`: best held-out checkpoint, selected at epoch 24 | |
| - Held-out Phase-1 PSNR: **34.862 dB** | |
| - Input resolution: 128³ volume, 25 sparse projections at 64² | |
| - X-ray fusion: strong fusion enabled | |
| - Null branch: disabled | |
| `config.json` records the exact training arguments. `train_log.txt`, | |
| `phase1_v3_training_curve.png`, and `phase1_v3_training_metrics.csv` provide | |
| the 30-epoch training record. | |
| ## Loading | |
| Use the V3 `phase_v27/export_prior_v27.py` exporter in this repository and | |
| pass `best.pth` as its checkpoint. The model architecture must be constructed | |
| with the matching `config.json` fields before loading the checkpoint. | |