Main Session
Sep
28
PQA 03 - Digital Health Innovation and Informatics, Patient Safety & Quality, and Radiation and Cancer Biology
Presenter(s)
Tyler Slater, MS - Case Western Reserve University School of Medicine, Cleveland, OH
T. Slater1, T. Arsenault2, K. O'Carroll2, S. K. Attia2, Y. Sun1, H. Newton3, C. Badve4, D. E. Spratt5, H. K. Perlow2, P. Vempati2, and A. Baydoun2; 1Case Western Reserve University School of Medicine, Cleveland, OH, 2Department of Radiation Oncology, University Hospitals Cleveland Medical Center/ Seidman Cancer Center, Cleveland, OH, 3Division of Neuro-Oncology, University Hospitals Cleveland Medical Center/ Seidman Cancer Center, Cleveland, OH, 4Department of Radiology, University Hospitals Cleveland Medical Center, Cleveland, OH, 5University Hospitals Seidman Cancer Center, Case Western Reserve University, Cleveland, OH
Purpose/Objective(s):
DOTATATE positron emission tomography (PET) has demonstrable clinical utility for intracranial meningioma, but adoption has been limited by infrastructure requirements and related costs. Artificial intelligence (AI)-based image-to-image translation could potentially address these limitations if comparable imaging information could be derived from magnetic resonance imaging (MRI). Here, we report a deep learning model that generates synthetic DOTATATE PET images from T1- and T2-weighted brain MRI.Materials/Methods:
Patients enrolled in a single-institution meningioma registry were included between 2021 and 2025. A conditional generative adversarial network was trained and evaluated using patient-level 10-fold cross validation. Input images were T1 post-contrast and T2 FLAIR MRI sequences. All images were resampled to a 1 mm isotropic resolution, co-registered, and padded or cropped to 256×256 pixels in the axial plane. A 2.5D approach was employed, incorporating adjacent slices to provide through-plane spatial context. In the evaluation phase, images were generated for the whole brain, and performance was quantified by comparing the synthetic and ground-truth PET images using mean absolute error (MAE), root mean squared error (RMSE), maximum standardized uptake value (SUV) error, and mean SUV error. Radiologist interpretation-derived tumor masks were used to compute tumor-specific metrics. Image quality was quantified using the structural similarity index (SSIM) and peak signal-to-noise ratio (PSNR).Results:
The dataset consisted of 40 patients (mean age 56.6 ± 18.5 [SD]; 9 males; 31 females) with 43 independent MRI-PET pairs due to repeated scans. Thirty patients underwent resection, with pathology demonstrating 8 WHO Grade I, 18 WHO Grade II, and 2 WHO Grade III lesions (2 patients had unknown WHO Grade). Average model performance metrics are summarized in Table 1. An MAE of 0.18 ± 0.08 SUV ([g/mL]) was achieved over the whole image, compared with 3.2 ± 2.1 SUV ([g/mL]) over the tumor area.Conclusion:
To our knowledge, this is the first reported model demonstrating the feasibility of AI-based generation of synthetic DOTATATE PET images from routine brain MRI. Future work includes refinement and external validation with multi-institutional data. If robustly validated, our AI-based synthetic DOTATATE PET could have global applications in the management of meningiomas. Table 1: Quantitative Performance Metrics for Synthetic versus Reference DOTATATE PET Images| Domain | Metric | Value (Mean ± SD) |
| Whole Image | MAE (SUV [g/mL]) | 0.18 ± 0.08 |
| RMSE (SUV [g/mL]) | 0.54 ± 0.20 | |
| Percent Bias (%) | 28 ± 47 | |
| SSIM (unitless) | 0.89 ± 0.04 | |
| PSNR (dB) | 34.8 ± 4.6 | |
| Tumor Only | MAE (SUV [g/mL]) | 3.2 ± 2.1 |
| RMSE (SUV [g/mL]) | 4.1 ± 3.0 | |
| Percent Bias (%) | -19 ± 45.7 | |
| SUVmax | Mean Error (SUV [g/mL]) | -7.6 ± 11.7 |
| Percent Bias (%) | -19 ± 71 | |
| SUVmean | Mean Error (SUV [g/mL]) | -2.5 ± 2.6 |
| Percent Bias (%) | -43 ± 46 |