Feasibility analysis of adrenal segmentation based on CT deep learning model

ZHANG Tuo; MENG Fanxing; PAN Yukun; KAN Xiaojing; GE Yinghui

doi:10.19300/j.2025.L21506

Abstract：Objective To explore the feasibility of training a deep learning model for fully automated adrenal segmentation on non-contrast CT images.Methods The images and clinical data of 1 200 patients who underwent non-contrast adrenal CT scan were retrospectively collected.Using a 5-fold cross-validation method,patients were divided into a training set(960 cases)and an internal test set(240 cases)at an 8∶2 ratio.Additionally,81 cases who underwent adrenal CT scans were collected as an independent test set.Both 2D nnU-Net and 3D nnU-Net segmentation models were constructed based on the nnU-Net framework.Clinical and CT imaging features were compared between the two groups using the Mann-Whitney U test and chi-square test.The model's segmentation performance was objectively evaluated using the Dice coefficient(DSC),Hausdorff distance(HD),average symmetric surface distance(ASSD),recall,and precision from the internal testing set and independent testing set.Two radiologists subjectively evaluated the 3D nnU-Net segmentation results on the independent test set.Results No statistically significant differences were observed in general characteristics between the training set+internal test set and independent test set(all P＞0.05).Both 2D and 3D nnU-Net models achieved high segmentation performance for the left and right adrenal glands on the internal and independent test sets.Compared to the 2D nnU-Net model,the 3D nnU-Net model demonstrated higher DSC and precision,lower HD and ASSD,and similar or higher recall.The segmentation results of the 3D nnU-Net were closer to manual annotations compared to the 2D nnU-Net model.Subjective evaluation by two radiologists on the independent test set revealed 62.96%satisfactory and 37.04%unsatisfactory segmentation outcomes for the 3D nnU-Net.Conclusion The deep learning-based adrenal segmentation mode is feasible for automatic adrenal segmentation on non-contrast CT images.