Towards Explainable and Automated PET/CT Interpretation: Combining AI-based CT-Segmentation with Rule-Based PET Analysis
Nuklearmedizin - NuclearMedicine · 01.04.2026 · 65 (02) · 123-124
DOI: 10.1055/s-0046-1818232 ↗Kurz gefasst
Fachartikel in der Zeitschrift Nuklearmedizin - NuclearMedicine, erschienen am 01.04.2026. Verfasst von M Galler und 6 weiteren Autor:innen.
Abstract (Originalsprache)
Ziel/Aim: Artificial intelligence (AI)-based methods have shown promising potential in enhancing PET/CT analysis. However, most existing approaches rely on end-to-end AI models, which suffer from limited explainability and high computational demands. Given the dual-modality nature of PET/CT imaging, it may be possible to divide the analysis workflow into distinct components and limit the use of AI to specific tasks. For example, AI could be employed solely for CT image segmentation, while subsequent PET data evaluation is performed using transparent, rule-based methods. Methodik/Methods: We propose an automatic PET/CT analysis workflow that restricts the use of AI to CT segmentation, leveraging existing tools for this purpose. As a proof of concept, we implemented this strategy for the fully automated detection and reporting of bone metastases in PSMA-PET/CT scans. The method involves AI-based bone segmentation using the nnU-Net–based tool TotalSegmentator. The resulting segmentations are applied to the PET scan, where each segment is evaluated for suspicious focal uptake using rule-based analysis, including maximum intensity detection and contrast-to-noise ratio assessment. Segmentation masks visually highlight detected regions. Segment-wise performance was evaluated on a test set of 100 [68Ga]Ga-PSMA-PET/CT scans, following parameter optimization on an independent training set of 20 scans. Ergebnisse/Results: The proposed method demonstrated strong diagnostic performance in a segment-wise evaluation of 6,099 bone segments. The algorithm achieved a sensitivity of 0.950 (CI: 0.925–0.966), a specificity of 0.968 (CI: 0.963–0.972), a PPV of 0.697 (CI: 0.659–0.733), and an NPV of 0.996 (CI: 0.994–0.997). Most misclassifications were attributed to inherent spatial resolution limitations of PET imaging or minor inaccuracies in anatomical segmentation. Schlussfolgerungen/Conclusion: By combining AI-based CT segmentation with rule-based PET analysis, this approach enables an interpretable and computationally efficient workflow for automated PET/CT interpretation. While promising, further development is required to extend the method to other organs and radiotracers and to improve segmentation accuracy. Publication History Article published online: 02 April 2026 © 2026. Thieme. All rights reserved. Georg Thieme Verlag KG Oswald-Hesse-Straße 50, 70469 Stuttgart, Germany
Zitieren
APA
Galler, M., Metzger, G., Schatka, I., Ruhwedel, T., Amthauer, H., Furth, C., & Rogasch, J. (2026). Towards Explainable and Automated PET/CT Interpretation: Combining AI-based CT-Segmentation with Rule-Based PET Analysis. Nuklearmedizin - NuclearMedicine, 65(02), 123-124. https://doi.org/10.1055/s-0046-1818232
Vancouver
Galler M, Metzger G, Schatka I, Ruhwedel T, Amthauer H, Furth C, et al. Towards Explainable and Automated PET/CT Interpretation: Combining AI-based CT-Segmentation with Rule-Based PET Analysis. Nuklearmedizin - NuclearMedicine. 2026;65(02):123-124. doi:10.1055/s-0046-1818232
BibTeX
@article{Galler2026,
author = {Galler, M and Metzger, G and Schatka, I and Ruhwedel, T and Amthauer, H and Furth, C and Rogasch, J},
title = {Towards Explainable and Automated PET/CT Interpretation: Combining AI-based CT-Segmentation with Rule-Based PET Analysis},
journal = {Nuklearmedizin - NuclearMedicine},
year = {2026},
volume = {65},
number = {02},
pages = {123--124},
doi = {10.1055/s-0046-1818232}
}
EUVIMED