Researchers have developed a weakly supervised framework named WiseSeg that completed whole-body PET/CT segmentation across 147 anatomical structures within 120 seconds per scan while achieving a mean Dice score of 0.874 on an independent test set. In comparative evaluations involving 113 structures matched to TotalSegmentator, the model outperformed TotalSegmentator with a mean Dice score of 0.895 compared to 0.884 (P < 0.001).
Whole-body organ segmentation is critical for radiotherapy planning and systemic disease assessment. However, existing PET/CT-tailored segmentation methods suffer from insufficient anatomical coverage, severe annotation bottlenecks, and an unfavorable trade-off between accuracy and speed, limiting their clinical translation.
To address these challenges, the retrospective study included 2,308 18F-FDG PET/CT scans from two hospitals and 1,746 public CT scans. Model development utilized 1,846 scans for pseudo-label training, consisting of 1,746 public CT scans and 100 PET/CT scans, alongside 2,000 unlabeled PET/CT scans for weakly supervised learning. Independent testing was conducted on 110 PET/CT scans, and a healthy cohort of 98 individuals was included to establish normal reference ranges.
The framework trained an uncertainty-guided pseudo-label learning algorithm to segment the 147 anatomical structures. Dice scores were calculated between model predictions and expert annotations to quantify accuracy, and paired t-tests or Wilcoxon tests with Bonferroni correction were used for statistical analysis.
The uncertainty-weighting strategy improved the segmentation of small vascular structures, reaching a Dice score of 0.702 compared to 0.651 (P < 0.001). In PET quantification, the model showed a lower absolute deviation in the 95th-percentile standardized uptake value (SUV95; 0.018 versus 0.021) than TotalSegmentator. Furthermore, normative SUV reference ranges for the 147 structures were established from the healthy cohort to provide a standardized benchmark for quantitative PET/CT interpretation, with source code and pretrained weights released publicly.
To address the limitations that hinder clinical translation—specifically insufficient anatomical coverage, severe annotation bottlenecks, and an unfavorable trade-off between segmentation accuracy and speed—the study formulated an uncertainty-guided pseudo-label learning algorithm. The resulting weakly supervised system was designed to delineate 147 anatomical structures across whole-body PET/CT imaging, supporting applications in radiotherapy planning and systemic disease assessment. By structuring model training around pseudo-labels derived from accessible CT data alongside targeted PET/CT scans, the methodology directly mitigates the requirement for exhaustive manual annotations across high-dimensional dual-modality volumes.
The retrospective research design utilized imaging resources collected from two hospital centers alongside public repositories, gathering a total of 2,308 18F-FDG PET/CT scans and 1,746 public CT scans. During model development, the researchers partitioned these data assets across specialized training stages. An initial pseudo-label training cohort combined the 1,746 public CT scans with 100 clinical PET/CT scans, providing 1,846 total imaging sets. Subsequently, 2,000 unlabeled PET/CT scans were incorporated into the weakly supervised learning phase, allowing the model to adapt to dual-modality properties without demanding manual segmentations. Evaluation was strictly isolated, utilizing 110 independent PET/CT scans reserved specifically for model testing.
To rigorously evaluate segmentation accuracy across complex anatomical geometry, the study benchmarked model predictions against expert annotations using Dice similarity coefficients. Statistical comparisons relied on paired t-tests or Wilcoxon tests, with Bonferroni correction applied to control for multiple hypothesis testing. On the independent 110-scan testing set, the architecture attained an overall mean Dice score of 0.874 across the complete panel of 147 structures, maintaining a processing speed of under 120 seconds per scan. For the 113 anatomical structures directly shared with the TotalSegmentator baseline, the framework demonstrated a statistically significant improvement in spatial overlap, recording a mean Dice score of 0.895 against 0.884 (P < 0.001).
The integration of an uncertainty-weighting strategy provided measurable performance advantages in challenging regions, specifically targeting small vascular structures where boundary definitions are often ambiguous. When evaluated on these vascular targets, the uncertainty-weighted model raised the mean Dice score from 0.651 to 0.702 (P < 0.001). Beyond structural boundary demarcation, the framework demonstrated enhanced fidelity in quantitative PET uptake analysis. In evaluations of metabolic activity, the model exhibited a lower absolute deviation in the 95th-percentile standardized uptake value (SUV95), registering 0.018 compared to 0.021 for TotalSegmentator, thereby reducing quantification discrepancies across segmented volumes.
In addition to algorithmic development and testing, the investigation established normative reference standards by analyzing a dedicated healthy cohort comprising 98 individuals. Using these healthy subjects, standardized uptake value distributions were mapped across all 147 anatomical structures to provide baseline reference ranges intended to support standardized, quantitative PET/CT interpretation. Published as peer-reviewed original research in EJNMMI Physics, the work provides an open-access resource, with source code and pretrained weights publicly available on GitHub to facilitate reproducibility and clinical translation.
Subject of Research: Weakly supervised whole-body PET/CT anatomical segmentation
Article Title: WiseSeg: weakly supervised whole-body PET/CT segmentation of 147 anatomical structures
Article References: Wen, P., He, J., Qiao, L., Chen, J., Tang, L., Li, A., Nie, Z., Gui, L., & Yang, X. (2026). WiseSeg: weakly supervised whole-body PET/CT segmentation of 147 anatomical structures. EJNMMI Physics. https://doi.org/10.1186/s40658-026-00934-9
Image Credits: AI Generated
DOI: 10.1186/s40658-026-00934-9
Keywords: PET/CT, image segmentation, weakly supervised learning, anatomical structures, TotalSegmentator, standardized uptake value, radiotherapy planning, pseudo-label learning, WiseSeg, weakly, supervised, whole-body
News Source: Denise Maddox. (November 18, 2026). Weakly Supervised Framework Segments 147 Anatomical Structures on Whole-Body PET/CT. Scienmag.



