TY - JOUR T1 - <strong>Artificial Neural Network for Prediction of Post-therapy Dosimetry for <sup>177</sup>Lu-PSMA I&amp;T Therapy</strong> JF - Journal of Nuclear Medicine JO - J Nucl Med SP - 1185 LP - 1185 VL - 60 IS - supplement 1 AU - Kuangyu Shi AU - Chao Dong AU - Andrei Gafita AU - Yu Zhao AU - Giles Tetteh AU - Bjoern Menze AU - Ali Afshar-Oromieh AU - Matthias Eiber AU - Axel Rominger Y1 - 2019/05/01 UR - http://jnm.snmjournals.org/content/60/supplement_1/1185.abstract N2 - 1185Purpose: The emerging PSMA targeted radionuclide therapy provides an effective method for the treatment of advanced metastatic prostate cancer. However, concerns of dose effects and risks have also been raised. The European council directive (council directive 2013/59 Euratom) mandates that treatments should be planned according to the radiation doses delivered to individual patients. However, there is no clinically practical method to predict the dosimetry before the internal radionuclide therapy, which hampers the realization of treatment planning. This study aims to prove the concept to employ artificial intelligence to predict the post-therapy dosimetry. Methods: A cohort of 43 patients with advanced metastatic prostate cancer were scanned with 68Ga-PSMA-11 PET/CT and then treated with 177Lu-PSMA I&amp;T. After the treatment, the patients underwent 3-5 planar whole-body scans for purpose of dosimetry. For proof-of-concept, we focus on organ-based prediction in this study. The mean SUV uptake of lung, kidney, brain, spleen, liver, heart, muscle, bladder and whole body were obtained from pretherapy PET/CT scans. Blood test values (PSA, Albumin etc) were also included as additional clinical information. Dosimetry were calculated for kidney, liver, spleen and salivary glands using Hermes Olinda 2. A 3-layer fully connected neural network was built up in Keras. The first two layers were processed with “ReLU” activate function. The network was optimized based on Mean Square Error with stochastic gradient descent method. 10-folder cross validation was applied to verify the trained network. The proposed individualized dosimetry prediction method was compared with population-based dosimetry from literature. The influence of blood sample information on prediction was also assessed. Results: The proposed artificial neural network achieved the dosimetry prediction error of 14.0±12.4% for kidney, 15.7±10.3% for liver, 77.5±12.8% for salivary glands and 24.3±16.1% for spleen. The inclusion of blood sample doesn’t reduce the prediction error (p&gt;0.9), 15.8±13.2% for kidney, 18.2±9.4% for liver, 72.2±15.8% for salivary glands and 28.1±19.9% for spleen. In contrast, the prediction based on literature population mean has significantly larger error (p&lt;0.01), 46.2±50.4% for kidney, 99.5±238.7% for liver, 705.7±377.7% for salivary glands. Conclusion: The proof of concept study shows that artificial neural network can significantly reduce the prediction error compared to generally population-based estimation. Artificial intelligence may provide a practical solution to improve the dosimetry-guided treatment planning for internal radionuclide therapy. 3D dosimetry including target lesions are currently under calculation to extensively test the proposed AI strategy. ER -