Enhancing MRI Diagnostics by Integrating QSM and Radiomics: A Feasibility Study on Age Effects

Background: QSM is effective in detecting pathological changes in tissue iron, myelin, and calcification. This approach may offer the potential to identify novel, non-invasive markers of brain health. Studying age-dependent trajectories of biomarkers is crucial for understanding brain development and aging, as well as predicting cognitive or neurodegenerative illnesses.
Aims: We aimed to assess the diagnostic potential of utilizing radiomics in QSM, reconstructed with a deep learning neural network, to characterize age-dependent effects. Additionally, we evaluated reproducibility by using two separate scans of the young cohorts.
Materials and Methods: A total of 48 participants were included in the study: 25 healthy older adults (mean age 50.28 ± 12.15 years) and 23 young subjects (mean age 20.74 ± 1.21 years). Imaging was conducted using 64-channel and 20-channel receiver array coils. All participants underwent multi-echo gradient echo data acquisition for QSM, which was reconstructed using a deep learning neural network. Radiomic parameters were extracted from QSM data through a systematic procedure encompassing four feature groups. Spearman correlation and the least absolute shrinkage and selection operator (LASSO) method were used for radiomics feature selection. A support vector machine (SVM) model was constructed to classify subjects from different age cohorts. The reproducibility of the QSM radiomics study was assessed by independently extracting radiomics features and scores from two scans of each young subject.
Results: QSM reconstruction using a deep learning neural network significantly enhanced prediction accuracy while reducing artifacts. Following feature selection with a significance level of p < 0.05, the feature set was reduced to 747 features. LASSO regression further refined the selection to 22 key features with high predictive value. These features were employed to train a SVM model, achieving an accuracy of 92.6%. The model demonstrated an Area Under the Curve (AUC) of 98%, with 93% sensitivity and 88.5% specificity. Additionally, radiomics scores were computed for each subject. Analysis of QSM radiomic scores revealed distinct differences between young and elderly subjects.
Conclusions: Our study highlights the feasibility of integrating radiomics into QSM as a promising diagnostic indicator. We established a statistically significant association between brain age and susceptibility scores. Our results further showed good reproducibility in radiomics scores when applying the selected features from one scan to another. This confirmed the reliability of combining QSM with radiomics in our study cohorts