Deep Adaptive Transfer Learning for Site-Specific PET Attenuation and Scatter Correction from Multi-National/Institutional Datasets
Bibliographic record
Abstract
Attenuation and scatter correction (ASC) must be performed for quantitative PET imaging. ASC is a challenging task in PET-only and PET/MRI systems. Different single-center or scanner-specific studies have been performed by using deep learning (DL) algorithms. However, the generalizability of these models is limited. In addition, providing large datasets for data-hungry DL models is a bottleneck. However, a universal model may not perform very well for each center because of high variability across different centers in scanners and image acquisition and reconstruction protocols. Therefore, we aimed to apply deep transfer learning for site-specific PET ASC utilizing a multi-center dataset. Altogether, 43068Ga-PSMA/DOTA PET/CT images from three countries (Switzerland, Iran, and Canada), including eight different centers, were enrolled in this study. In all centers, PET images were corrected by using CT images for ASC. In addition, we implemented a deep supervised U-Net network architecture, U2Net, as the core DL model. Different scenarios were investigated in this study, including (i) training and testing models for each center separately: center-based (CeBa); (ii) training and testing models using the entirety of the dataset: centralized (CeZe); i.e., entire data pooled together; and (iii) transfer learning (TrLe) where training was performed using pooled data to one server using entire dataset and then TrLe were performed for each center separately to build site-specific models for each center. In terms of absolute relative error (ARE%), CeBa, CeZe, and TrLe achieved 36 ± 13 (CI95%: 33 to 39), 31 ± 34 (CI95%: 24 to 38).Furthermore, using TrLe outperformed the centralized and center-based models in terms of accurate ASC image generation.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.003 | 0.005 |
| Meta-epidemiology (narrow) | 0.002 | 0.001 |
| Meta-epidemiology (broad) | 0.001 | 0.001 |
| Bibliometrics | 0.001 | 0.002 |
| Science and technology studies | 0.001 | 0.001 |
| Scholarly communication | 0.001 | 0.001 |
| Open science | 0.002 | 0.002 |
| Research integrity | 0.001 | 0.002 |
| Insufficient payload (model declined to judge) | 0.001 | 0.001 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".