Predicting 18F-FDG SUVs of metastatic pulmonary nodes from CT images in patients with differentiated thyroid cancer by using a convolutional neural network

Purpose: The aim of this study was to predict standard uptake values (SUVs) from computed tomography (CT) images of patients with lung metastases from differentiated thyroid cancer (DTC-LM).

Methods: We proposed a novel SUVs prediction model using 18-layer Residual Network for generating SUVmax, SUVmean, SUVmin of metastatic pulmonary nodes from CT images of patients with DTC-LM. Nuclear medicine specialists outlined the metastatic pulmonary as primary set. The best model parameters were obtained after five-fold cross-validation on the training and validation set, further evaluated in independent test set. Mean absolute error (MAE), mean squared error (MSE), and mean relative error (MRE) were used to assess the performance of regression task. Specificity, sensitivity, F1 score, positive predictive value, negative predictive value and accuracy were used for classification task. The correlation between predicted and actual SUVs was analyzed.

Results: A total of 3407 nodes from 74 patients with DTC-LM were collected in this study. On the independent test set, the average MAE, MSE and MRE was 0.3843, 1.0133, 0.3491 respectively, and the accuracy was 88.26%. Our proposed model achieved high metric scores (MAE=0.3843, MSE=1.0113, MRE=34.91%) compared with other backbones. The predicted SUVmax (R2 = 0.8987), SUVmean (R2 = 0.8346), SUVmin (R2 = 0.7373) were all significantly correlated with actual SUVs.

Conclusion: The novel approach proposed in this study provides new ideas for the application of predicting SUVs for metastatic pulmonary nodes in DTC patients.