TY - JOUR T1 - Deep Supervised Residual U-Net for Automatic Characterization of Lesions on<sup>68</sup>Ga-PSMA PET/CT images JF - Journal of Nuclear Medicine JO - J Nucl Med SP - 1217 LP - 1217 VL - 60 IS - supplement 1 AU - Yu Zhao AU - Andrei Gafita AU - Giles Tetteh AU - Fabian Haupt AU - Ali Afshar-Oromieh AU - Bjoern Menze AU - Matthias Eiber AU - Axel Rominger AU - Kuangyu Shi Y1 - 2019/05/01 UR - http://jnm.snmjournals.org/content/60/supplement_1/1217.abstract N2 - 1217Purpose: The emerging PSMA targeted radionuclide therapy provides an effective method for the treatment of advanced metastatic prostate cancer. To optimize the therapeutic effect and maximize the theranostic benefit, there is an urgent need to identify and quantify target lesions prior to treatment. However, this is extremely challenging considering that a high number of lesions of heterogeneous size and uptake may distribute in a variety of anatomical context with different backgrounds. Until now, there are no successful computer-aided lesion detection methods for PSMA-ligand PET imaging. This study proposes an automated lesion detection method for prostate cancer (PC) via a specially designed deep supervised residual U-Net to facilitates clinal practice. Methods: A dataset included 71 patients (mean age 71.0 ± 8.4 years, range 51-85 years) with metastatic prostate cancer was collected from three medical centers. These patients underwent 68Ga-PSMA-11 PET/CT imaging from the head to the thigh. For proof-of-concept, we focus on the detection of bone &amp; lymph node lesions in the pelvic area. The bone and lymph node lesions were manually labelled by a nuclear medicine expert. A 3D deep supervised residual U-Net was developed to detect the lesions. It works by first extracting salient features from PET and CT and then adopting the combined features to automatically detect all the lesions in a 3D manner. The network comprises a down-sampling path including four repeated encoder stacks and an up-sampling path including four repeated decoder stacks. The down-sampling path aggregates increasingly abstract information and the up-sampling path then recombines this information with shallower features to precisely localize the structures of the interest. Compared to traditional 3D U-Net, we employed deep supervision, residual connection and instance normalization to improve the performance of the neural network. Results: The proposed method achieves precision of 97% ± 9%, recall of 98% ± 7%, F1 score of 98% ± 8% on bone lesion detection and precision of 75% ± 35%, recall of 84% ± 21%, F1 score of 79% ± 22% on lymph node lesion detection, which demonstrates the effectiveness of the proposed method on PC lesion detection. The segmentation accuracy is relatively low compared to the detection accuracy with average dice score for bone and lymph node of 0.75 and 0.40, respectively. The figure demonstrates example segmentation of the proposed network. Notably, the performance of the network was influenced by the training data, in particular for lymph node lesions where only limited training samples were available. Conclusions: This study proposed an end-to-end deep neural network to detect and segment the PC lesions on PSMA imaging automatically. The preliminary test on pelvic area confirms the potential of deep learning methods. Increasing the amount of training data may further enhance the performance of the proposed deep learning method. Currently, more data are in collection and annotation. Increasing the amount of training data may further enhance the performance of the developed deep learning methods. ER -