Novel AI-Based Dose Prediction Directly from Diagnostic PET/CT: Applications for Multi-Disciplinary Lung Cancer Care

Purpose: Multi-disciplinary clinics are becoming standard in oncology. They benefit from early prediction of doses to organs-at-risk (OAR) for improved clinical decision support. However, this information is not available until a treatment plan is made. This work develops an AI-based dose prediction algorithm that only uses diagnostic PET/CT scans for lung cancer care, providing crucial dose prediction before treatment planning.
Methods: The study involved 75 lung cancer patients from two institutions who had both diagnostic PET/CT scans and dose information available. The dose distribution was rigidly registered to the PET/CT to evaluate the prediction results of 3D CNN and Transformer models (3D UNET, 3D Attention UNET, UNETR, and SwinUNETR). Models were trained with different input combinations: PET/CT, PET only, and CT only. Predicted dose distributions in the heart were compared with clinical dose distribution using root mean square error (RMSE). Mean heart dose differences (MHDD) were evaluated to assess the clinical utility of dose prediction.

Results: The models achieved high accuracy in predicting dose distributions. Notably, PET/CT inputs consistently outperformed CTs, and SwinUNETR outperformed other models except UNETR with PET/CT.

Conclusion: The study demonstrated that AI models can help determine diagnostic PET/CT image dose distributions. The impact of this work lies in the availability of MHD predictions using diagnostic imaging alone, facilitating improved decision support in multi-disciplinary settings where radiation treatment plan data is not yet available.

1. Novel AI-based Dose prediction directly from Diagnostic PET/CT: applications for multi-disciplinary lung cancer care Wookjin Choi, PhD , Yingcui Jia, PhD, Wentao Wang, PhD and Yevgeniy Vinogradskiy, PhD, Thomas Jefferson University, Philadelphia, PA

2. Supported by NIH/NCI R01 CA236857 and the NIH/NCI Cancer Center Support Grant 5P30 CA056036. Research grant from Varian Medical Systems, Inc. Acknowledgements & Disclosure

3. Introduction Dataset Methods Results Conclusions and Future works Contents 3

4. Multi-disciplinary clinics are becoming standard in oncology. Early prediction of doses to organs-at-risk (OAR) improves clinical decision support. This information is typically unavailable until a treatment plan is made. Develop an AI-based dose prediction algorithm using diagnostic PET/CT scans for crucial dose prediction before treatment planning. Heart toxicity: Mean Heart Dose > 5Gy 4 Introduction Diagnostic PET/CT Radiation dose distribution Typical Radiation Oncology workflow 2 Weeks Contouring Planning CT Sim

5. Proposed solution 5 Diagnostic PET/CT Radiation dose distribution 2 Weeks Contouring Planning CT Sim A few minutes AI model

6. Dataset The study involved 75 lung cancer patients from two institutions (TJU: 36, CU: 39) Diagnostic PET/CT scans Dose information The dose distribution was rigidly registered to the PET/CT Image dimensions were resampled to match dose distribution (2.5mmx2.5mmx3mm) 6

7. Methods 3D Convolutional Neural Network (CNN) and 3D Transformer models UNET Attention UNET UNETR SwinUNETR The models utilized input patches of 96x96x96. Input: PET/CT, PET only, or CT only Output: Dose map Training 80% (n=60) and Testing 20% (n=15) Model trained on NVIDIA RTX A6000 48GB Data augmentation: random sampling of cropped patches, random flips and 90 degree rotation on x, y, z axis, random intensity shift) Root Mean Square Error (RMSE) loss between clinical dose and predicted dose CNN models: 10 samples, 8 batches, 1000 epochs Transformer models: 5 samples, 2 batches, 1200 epochs Evaluation of predicted doses RMSE of Heart Volume Mean Heart Dose Difference (MHDD) 7

8. Results Model PETCT PET CT UNET 6.0±2.9 5.7±3.2 7.8±3.6 Attention UNET 4.9±2.3 5.8±3.2 7.8±3.2 UNETR 2.3±1.7 2.4±1.7 2.6±1.7 SwinUNETR 2.7±1.7 2.3±1.3 2.9±1.8 RMSE of Heart Volume ( Gy ) 8 Mean Heart Dose Difference (MHDD) ( Gy )

9. Mean heart dose: 7.1 Gy Max cord dose: 34.2 Gy Lung_L V20: 42% Mean heart dose: 9.9 Gy Max cord dose: 39.6 Gy Lung_L V20: 47% 9 External Sample Evaluation

10. We also collected 113 lung SBRT cases Preliminary results UNET and Attention UNET accurately predict the dose distribution. The UNET PET-only model had larger values for both metrics, suggesting the necessity of using PET/CT input while using UNET models to provide both functional and anatomical information Results: Lung SBRT Model Input RMSE MHDD UNET PET/CT 2.91 ± 4.86 0.69 ± 3.81 CT 2.84 ± 5.07 1.11 ± 3.93 PET 7.89 ± 3.86 -5.04 ± 4.23 Attention UNET PET/CT 2.45 ± 4.86 0.90 ± 3.75 CT 2.64 ± 5.09 0.84 ± 3.92 PET 2.83 ± 4.73 0.64 ± 3.80 10 Mean Heart Dose Difference (MHDD) ( Gy )

11. AI models accurately predict dose distributions using diagnostic PET/CT. PET/CT input is essential for improved prediction accuracy. This approach enhances decision support in multi-disciplinary clinics. Provides early OAR dose information using only diagnostic PET/CT Facilitates improved decision-making in multi-disciplinary clinics Add more patient data to improve model robustness. Current total N=264, TJU: 3D-112, SBRT-113, CU: 3D-39 Continue to refine and enhance AI models. Provide additional DVH metrics Incorporate OARs as inputs using MIM AI Contour for quick delineations. Incorporate with Cardiac Event prediction model Conclusions and Future Works 11

12. Thank you! 12

13. Jefferson.edu