Hippocampal auto-segmentation based on deep learning for identifying magnetic resonance imaging biomarkers of early...

The study developed a predictive model based on deep learning for automatic segmentation of hippocampal structure, along with radiomic feature extraction, and built various machine learning diagnostic models for mild cognitive impairment (MCI).

The study cohort of 150 subjects was randomly partitioned into training and validation datasets at a 7:3 ratio. The hippocampus was automatically segmented by a new deep-learning network based on convolutional neural network (CNN) model in 3D T1WI. A total of 1183 radiomic features were systematically extracted from the bilateral hippocampus of each subject. Four independent machine learning (ML) classifiers, including logistic regression (LR), support vector machine (SVM), random forest (RF), and XGBoost algorithm were trained on the training set, and validated on the testing set in the form of 5-fold cross-validation.

Comprehensive evaluation of multiple ML classifiers, identified XGBoost as the most effective algorithm (AUC = 0.935, 95 %CI:0.83-0.97) in discriminating normal controls and MCI. The XGBoost model, utilizing the seven selected features, achieved the highest AUC in the test set (AUC = 0.864, 95 %CI:0.73-0.89). The glszm_ZoneVariance feature demonstrated the highest predictive weight among the radiomic features incorporated in the XGBoost model. Optimal net benefit was observed at threshold probabilities of 0.18 and 0.17 for patients with MCI in training and test sets.

A predictive model leveraging deep learning for automatic hippocampal segmentation, combined with radiomic features integrated into an XGBoost framework, may offer a promising biomarker for diagnosing MCI.