摘要Objective:To construct a prediction model for three-dimensional(3D)dose distribution of iodine-131 based on SPECT/CT radiomics features.Methods:A multi-scale feature pyramid network(MSFPN)was used to extract heterogeneity features of thyroid tissues before and after iodine-131 treatment.A spatiotemporal modeling framework(MSFPN+CNN+GAT+3D model)integrating convolutional neural network(CNN)and graph attention network(GAT)was established to achieve precise dose distri-bution prediction.Clinical and imaging data from 320 patients at a tertiary hospital were divided into training,validation,and test sets at a 7∶2∶1 ratio.The models were evaluated for iodine-131 residence time and time-activity curves(TACs)of key target organs.Pre-dictive accuracy was assessed using root mean square error(RMSE),mean absolute error(MAE),and gamma index.Results:The residence time of iodine-131 in the thyroid,bladder,and stomach was longer than that in the bone marrow(P<0.05).Following io-dine-131 treatment,the activity curve of bone marrow showed minimal variation over time.The bladder tissue exhibited an initial increase in activity,reaching its peak at 4 h,followed by a gradual decline.Both the thyroid and gastric tissues demonstrated a de-creasing trend in activity over time,with the gastric tissue displaying even lower dose levels compared to the thyroid.The RMSE and MAE of the MSFPN+CNN+GAT+3D model for the dose distribution of each target organ were lower than those of the MSFPN,ResNet and 3D CNN models,and the γ index was higher than those of the MSFPN,ResNet and 3D CNN models(P<0.05),and there was no statistical significance when compared with the ensemble model(P>0.05).Conclusion:The MSFPN+CNN+GAT+3D model systematically captures deep radiomics features,effectively facilitating the imag-ing evaluation of the accurate dose distribution of multiple organs during radioactive io-dine-131 treatment.