Deep Learning Automates PET RANO Criteria Extraction in Glioma Imaging

Standardizing Metabolic Assessment in Glioma Management

Managing primary brain tumors requires precise imaging to differentiate between active disease and treatment-related changes, a task where conventional magnetic resonance imaging often falls short [1, 2]. Amino acid tracers like [3]fluorodopa (FDOPA) have emerged as essential adjuncts, offering high sensitivity for detecting tumor recurrence and identifying high-grade transformation before contrast enhancement appears on standard imaging [4, 5]. While these positron emission tomography (PET) tracers provide critical quantitative metrics such as metabolic tumor volume and tumor-to-background ratios, the manual extraction of these parameters remains technically demanding and subject to significant inter-institutional variability [6, 7]. Current clinical guidelines emphasize the need for standardized protocols to harmonize data acquisition and interpretation across oncology centers [8, 9]. To address these logistical hurdles, researchers have developed and validated an automated deep learning tool to quantify metabolic tumor burden across multiple clinical sites, potentially streamlining how neuro-oncologists monitor treatment response.

Multicenter Development of a Volumetric Segmentation Model

To address the need for standardized metabolic assessment, researchers developed a deep learning model designed for the automated extraction of PET Response Assessment in Neuro-Oncology (RANO) criteria from [18F]FDOPA PET scans. The RANO criteria provide the standard clinical framework for evaluating whether a brain tumor is progressing or responding to therapy. This study utilized a robust multicenter dataset comprising 635 static [18F]FDOPA PET scans retrospectively collected from three European medical centers. These scans represented a broad clinical spectrum, including patients undergoing initial glioma diagnosis, those being evaluated for potential recurrence, and individuals requiring longitudinal treatment monitoring. By including diverse clinical indications, the authors aimed to ensure the model could handle the varied metabolic presentations encountered in daily neuro-oncological practice. The dataset was strategically partitioned across the participating institutions to ensure rigorous training and objective evaluation. The training cohort consisted of 530 scans from Nancy Hospital, providing a substantial foundation for the algorithm to learn complex metabolic patterns. For external validation, the researchers utilized 66 scans from Pitié-Salpêtrière Hospital, while an additional 39 scans from Turin Hospital served as an independent test set. This multicenter approach is critical for clinicians, as it tests whether the software can maintain accuracy across different scanners and institutional imaging protocols. The technical backbone of the system is a 3D U-Net, a type of artificial neural network specifically optimized for processing three-dimensional medical images and volumetric segmentation. This model was trained to segment both the tumor and healthy brain volumes, allowing for the calculation of relative metabolic metrics. To ensure the highest level of accuracy, the researchers established ground truth segmentations by following international guidelines, which served as the gold standard for training the algorithm. By automating the identification of tumor boundaries, the 3D U-Net aims to eliminate the subjectivity and time constraints associated with manual metabolic tumor volume quantification.

High Concordance with Expert Manual Segmentation

The researchers evaluated the model's performance using the Dice coefficient, a statistical metric ranging from 0 to 1 that measures the spatial overlap between the automated segmentation and the expert ground truth. This metric helps clinicians understand how closely the software mimics the manual contouring performed by experienced nuclear medicine physicians. In the training cohort, the tumor segmentation achieved a Dice coefficient of 0.925 [0.841; 0.970], indicating a high degree of spatial overlap. This level of accuracy remained robust as the model was applied to external data, which is a critical test of its reliability across different clinical environments. The model achieved a Dice coefficient of 0.885 [0.829; 0.925] in the validation set and maintained a Dice coefficient of 0.851 [0.733; 0.911] in the independent test set. Beyond volumetric overlap, the clinical utility of the tool depends on its ability to accurately detect the presence of disease. The study found that measurable lesions were correctly identified by the model in more than 97% of cases, suggesting the algorithm is highly sensitive for identifying tumor foci that require clinical monitoring. For the practicing physician, these results indicate that the deep learning model provides a reliable, automated alternative to manual segmentation. This capability could significantly reduce the inter-observer variability that often complicates the longitudinal assessment of gliomas using amino acid PET imaging, ensuring that a reported change in tumor volume reflects true biology rather than a difference in human measurement.

Reliability of Automated Metabolic Metrics

The clinical utility of the deep learning model depends on its ability to accurately replicate the quantitative metrics used in routine neuro-oncology practice. The researchers assessed the quantitative agreement for PET RANO criteria 1.0 parameters, which are the standardized metrics used to evaluate treatment response in gliomas. These parameters include the metabolic tumor volume (MTV), which measures the total volume of the tumor exhibiting increased tracer uptake, and tumor-to-background ratios, which compare the intensity of tracer uptake in the lesion to that of healthy brain tissue. Specifically, the study evaluated the tumor-to-background ratio maximum (TBRmax) and the tumor-to-background ratio mean (TBRmean). When compared to expert quantification at the lesion level, the automated model demonstrated high precision and low bias across all metabolic metrics. The agreement for metabolic tumor volume showed a bias of 2.293 [-4.734; 9.321] mL, indicating that the software's volumetric measurements closely align with manual expert contouring. For intensity-based metrics, the tumor-to-background ratio maximum showed a bias of 0.056 [-0.189; 0.301], while the tumor-to-background ratio mean showed a bias of -0.139 [-0.424; 0.146]. These narrow ranges of bias suggest that the automated tool can reliably substitute for manual calculations without introducing significant errors in the assessment of tumor metabolic activity. The consistency of these findings is further supported by intra-class correlation coefficients superior to 0.93 for all parameters. The intra-class correlation coefficient is a statistical index used to measure the reliability of measurements made by different observers, where a value exceeding 0.90 represents excellent reliability. For the practicing clinician, this high level of agreement suggests that the automated system provides results as dependable as those produced by experienced nuclear medicine physicians. To support broader clinical adoption and further research, the researchers have made the deep learning model publicly available via the IADI-Nancy GitHub repository, providing a tool for fully automated and reproducible quantification of glioma metabolism.

References

1. Somme F, Bender L, Namer IJ, Noël G, Bund C. Usefulness of 18F-FDOPA PET for the management of primary brain tumors: a systematic review of the literature. Cancer Imaging. 2020. doi:10.1186/s40644-020-00348-5

2. Santo G, Laudicella R, Linguanti F, et al. The Utility of Conventional Amino Acid PET Radiotracers in the Evaluation of Glioma Recurrence also in Comparison with MRI. Diagnostics. 2022. doi:10.3390/diagnostics12040844

3. Pang Y, Gupta G, Jha A, et al. Nonmosaic somatic HIF2A mutations associated with late onset polycythemia‐paraganglioma syndrome: Newly recognized subclass of polycythemia‐paraganglioma syndrome. Cancer. 2019. doi:10.1002/cncr.31839

4. Roach JR, Plaha P, McGowan DR, Higgins GS. The role of [18F]fluorodopa positron emission tomography in grading of gliomas. Journal of Neuro-Oncology. 2022. doi:10.1007/s11060-022-04177-3

5. Sharma A, Garg A, Singh M, et al. Metabolic imaging in recurrent gliomas: comparative performance of 18F-FDOPA, 18F-fluorocholine and 18F-FDG PET/CT. Nuclear Medicine Communications. 2023. doi:10.1097/mnm.0000000000001795

6. Xiao J, Jin Y, Nie J, Chen F, Ma X. Diagnostic and grading accuracy of 18F-FDOPA PET and PET/CT in patients with gliomas: a systematic review and meta-analysis. BMC Cancer. 2019. doi:10.1186/s12885-019-5938-0

7. Piñeiro-Fiel M, Moscoso A, Pubul V, Ruibal Á, Silva‐Rodríguez J, Aguiar P. A Systematic Review of PET Textural Analysis and Radiomics in Cancer. Diagnostics. 2021. doi:10.3390/diagnostics11020380

8. Verger A, Tolboom N, Cicone F, et al. Joint EANM/EANO/RANO/SNMMI practice guideline/procedure standard for PET imaging of brain metastases: version 1.0. European Journal of Nuclear Medicine and Molecular Imaging. 2025. doi:10.1007/s00259-024-07038-5

9. Amro S, Vinjamuri S, Dizdarevic S. PET computed tomography policy variations across the UK: a call for harmonisation and equity of access. Nuclear Medicine Communications. 2026. doi:10.1097/mnm.0000000000002107