TY - JOUR T1 - AI-based quantification of PET/CT lesions is associated with survival in lung cancer patients. JF - Journal of Nuclear Medicine JO - J Nucl Med SP - 1042 LP - 1042 VL - 62 IS - supplement 1 AU - Pablo Borrelli AU - Jose Luis Loaiza Gongora AU - Reza Kaboteh AU - Johannes Ulen AU - Olof Enqvist AU - Elin Tragardh AU - Lars Edenbrandt Y1 - 2021/05/01 UR - http://jnm.snmjournals.org/content/62/supplement_1/1042.abstract N2 - 1042Introduction: Total lesion glycolysis (TLG) measurements derived from 18-F-fluorodeoxyglucose (FDG) PET/CT studies have been shown to be a prognostic factor for overall survival (OS) in patients with lung cancer. An objective method to measure TLG is needed to compare results from different centers. The aim of this study is to develop an artificial intelligence (AI) tool for detection and quantification of lung tumors and thoracic lymph nodes in PET/CT studies and to evaluate the association between the TLG measurements and OS. Methods: The training group consists of 214 patients referred for FDG PET/CT due to suspected lung cancer. The separate test group consists of 108 consecutive patients referred for FDG PET/CT due to suspected lung cancer at Skåne University Hospital, Lund and Malmö, Sweden during 2018. Patients were injected with 4 MBq/kg 18-F-FDG and after 60 min scanned on a Discovery MI PET/CT (GE Healthcare, Milwaukee, WI). Two nuclear medicine specialists with >6 and >12 years of PET/CT experience segmented lung tumors and thoracic lymph nodes in the PET/CT studies of the training and test groups.The model consists of the organ convolutional neural network (CNN) from the work by Trägårdh et al. and a novel detection CNN trained to find lung tumors using three inputs; CT, PET and a mask from the organ CNN with one channel each for bone, liver, lung, heart, aorta and adrenal gland. A 3D U-Net detection CNN was trained using patches, half chosen from the background and half random patches with lung tumor pixels. Training was performed for 100 epochs using 10,000 patches per epoch. After this phase, the model was applied to the training set. The model was then retrained with 20% of the patches focusing on pixels incorrectly classified by the older model. These steps were repeated four times to produce the final model.The output from the detection CNN was thresholded to produce a tumor mask and connected components with a TLG below 0.1 was removed.Associations between TLG measurements and OS were studied using a univariate Cox-regression model. Results: A total of 51 patients died during the follow-up period, with a median survival time from the PET/CT study of 0.9 years (IQR 0.55-1.54). The group of 57 patients that were still alive had a median follow-up time from the PET/CT study of 2.6 years (IQR 2.5-2.7).There were good correlations between manual and AI-based TLG measurements for lung tumors (r=0.97) and lymph node lesions (r=0.79). The AI-based TLGmeasurements for lung tumor (p=0.004) and lymph node lesions (p=0.01) were significantly associated with OS. The corresponding manual measurements for lung tumor (p=0.002) and lymph node lesions (p=0.04) were also significantly associated with OS.The AI tool detected lung tumors in 94/98 (96%) manually positive patients and no lung tumors in 5/10 (50%) manually negative patients. The corresponding values for lymph nodes were 60/60 (100%) positive patients and 12/48 (25%) negative patients. A total of 161 false positive lung tumors (1.5 per patient) and 284 false positive lymph nodes (2.6 per patient) were detected by the AI tool in the 108 patients. Conclusions: Fully automated AI-based measurements showed prognostically significant association with OS. This type of measurements may be of value in the management of future patients with lung cancer. The AI tool developed in this project is available upon reasonable request for research purposes at www.recomia.org. References: Trägårdh E, et al. RECOMIA-a cloud-based platform for artificial intelligence research in nuclear medicine and radiology. EJNMMI Phys. 2020;7:51. ER -