PBPK Pre-trained Deep Learning for Voxel-wise Prediction of Post-therapy Dosimetry for 177Lu-PSMA Therapy

1174 Introduction: PSMA-directed radioligand therapy (RLT) has become one of the effective treatment options for metastatic castration-resistant prostate cancer (mCRPC). In SNMMI 2020, we proposed a deep learning method for voxel-wise prediction of post-therapy dosimetry from pre-therapy positron emission tomography (PET) images. However, the accuracy is still less satisfactory due to limited data. In this study, we propose to integrate physiologically based pharmacokinetic (PBPK) model in the pretraining of the deep learning methods to improve the prediction. Methods: 23 patients with mCRPC treated with 177Lu-PSMA I&T RLT and 11 patients treated with 177Lu-PSMA-617 were retrospectively included in this study. Only those cycles with pre-therapy PET imaging before the treatment and at least 3 post-therapeutic SPECT/CT dosimetry imaging were selected. Totally 48 treatment cycles from 177Lu-PSMA I&T and 11 cycles from 177Lu-PSMA-617 were considered for this proof-of-concept study. 3D RLT DoseGAN were developed with a 3D U-net generator and a convolutional neural network (CNN) based discriminator. For pretraining, 266 digital phantoms were generated based on the PBPK modeling on XCAT phantoms to simulate a variety of pretherapy PET and the spatiotemporal distribution of therapy ligands. A forward projection PET simulator was employed to simulate PET imaging on digital phantoms and dose-kernel methods were applied to generate post-therapy dosimetry. Results: The preliminary results showed that, the PBPK pre-trained 3D RLT Dose GANs achieved the voxel-wise normalized root mean squared error (NRMSE) of 3.2±0.7% (mean±std.) (3.8±0.7% without pre-training) and peak signal-to-noise ratio (PSNR) of 30.1±1.8 (28.5±1.6% without pre-training) on 177Lu-PSMA I&T. As for PSMA-617, it achieves 2.1±0.8% NRMSE (2.2±0.7% without pre-training) and PSNR of 34.0±3.8 (34.0±3.9% without pre-training) on 18F dataset, and 1.8±0.9% NRMSE (1.9±0.8% without pre-training) and PSNR of 35.0±2.5 (34.5±3.9% without pre-training) on 68Ga dataset. Furthermore, clinical assessment with coronal maximum intensity projection (MIP) and dose volume histogram (DVH) also confirmed that our proposed model achieved similar performance on examples from these two datasets in terms of image quality. The AI-generated dosimetry images achieved a general mean absolute error (MAE) of 21.2±10.8% (24.0±10.0% without pre-training) to the ground truth, in terms of DVH. Conclusions: Our experimental results demonstrate the incorporation of PBPK model may improve the development of artificial intelligence methods for dosimetry prediction and accelerate the implementation of dosimetry-guided treatment planning for RLT.