Automated liver lesion detection in 68Ga DOTATATE PET/CT using a deep fully convolutional neural network

Abstract Background Gastroenteropancreatic neuroendocrine tumors most commonly metastasize to the liver; however, high normal background 68 Ga-DOTATATE activity and image noise make metastatic lesions difficult detect. The purpose of this study is develop a rapid, automated highly specific method identify PET/CT hepatic using 2D U-Net convolutional neural network. Methods A retrospective patient studies ( n = 125; 57 with without) was evaluated. dataset randomly divided into 75 for training set (36 abnormal, 39 normal), 25 validation (11 14 normal) testing normal). Hepatic were physician annotated modified PERCIST threshold, boundary definition by gradient edge detection. trained independently five times 100,000 iterations linear combination binary cross-entropy dice losses stochastic descent algorithm. Performance metrics included: positive predictive value (PPV), sensitivity, F 1 score area under precision–recall curve (PR-AUC). Five different pixel thresholds used filter noisy predictions. Results total 233 each abnormal containing mean 4 ± 2.75 lesions. 20 produced highest PPV 0.94 0.01. 5 sensitivity 0.74 0.02. 0.79 0.01 filter. PR-AUC 0.73 0.03 15 Conclusion Deep networks can automatically detect in PET. Ongoing improvements data annotation methods, increasing sample sizes methods are anticipated further improve detection performance.