Introduction

Chemical exchange saturation transfer (CEST) imaging is an emerging MRI technique that can detect proteins and metabolites in vivo (1,2). Amide proton transfer (APT) imaging is a subtype of the CEST technique, which measures a bulk water signal reduction at 3.5ppm downfield due to its chemical exchange with amide protons (3). Studies have shown that the average APT-weighted (APTw) image intensities acquired from high-grade tumors were significantly higher than those from low-grade tumors, demonstrating the clinical potential of APT MRI (3,4). Radiomics is the process of converting regular medical images into mineable high-dimensional data (5), extracting semantic and/or agnostic features that may not be perceivable by human eyes. In this work, we implemented a radiomic analysis of APTw images and demonstrated its advantage for discriminating high-grade and low-grade brain tumors.

Methods

Data Acquisition: Forty pediatric patients (5.7 ± 4.4 years old) were enrolled in the experiments, including 20 cases with confirmed high-grade brain tumors and 20 with low-grade brain tumors. All the experiments were performed on a 3T Philips Achieva scanner with the following scanning parameters: RF saturation power/duration = 2uT/0.8 sec, TR/TE = 3000/6.7 ms, FOV = 186 x 230 mm², slice thickness = 6mm, 63 frequency offsets from -6 to 80ppm, and total acquisition duration = 3.2 min. The APTw image was calculated by subtracting the reference signal at -3.5ppm from the label signal at +3.5ppm with MTR asymmetry (MTRasym) analysis (6).
Semi-automatic Habitat Definition: The region of interest (ROI) encircling the whole tumor was initially drawn by an experienced radiologist. Then, an automatic ROI-shrinking algorithm was used to choose sub-regions with signals greater than a histogram threshold (7). In our experiments, we generated four sets of ROIs (using 0^th, 20^th, 50^th, and 80^th percentile of the histogram as the cutoff), and calculated the mean APTw signals within the ROIs.
Radiomics Feature Extraction: For each tumor volume inside the automatically-generated ROI, a total number of 465 features were extracted using Pyradiomics (8). The feature set included 10 two-dimensional shape features, 18 first-order statistical features, 22 GLCM features, 16 GLRLM features, 16 GLSZM features, 14 GLDM features, and 5 NGTDM features. The first-order statistical features and the texture features were extracted from the original images and four wavelet-transformed images. The values of each feature were mapped to 0 – 1 by Min-Max normalization.
Feature Selection and Modelling: A feature selection method with 5-fold stratified cross-validation was implemented. A recursive feature elimination (RFE) method was implemented on the training set by fitting a logistic regression model, removing the most irrelevant feature, and assessing the performance of the remaining feature set recursively. Only the three top-ranked features were selected at the end of each RFE process, and the average APTw signal was added as a mandatory fourth feature in order to improve the generalizability of the model. Based on the selected features, we trained support vector machine (SVM) classifiers for grading tumors. The parameters of the SVM classifiers (including the kernel function, kernel function parameters, and regularization parameter) were optimized by a grid search to achieve the best performance on the training set. The trained model was then evaluated on the test set, with the average sensitivity, specificity, and area under the receiver operating characteristic curve (AUC) calculated. The process of feature selection with 5-fold stratified cross-validation was repeated 100 times with the 5-fold data partition randomly shuffled, and the pooled results were statistically tested with IBM SPSS Statistics v27. The best feature set was identified according to the frequency distribution of top-ranked features generated during feature selection (9).

Results

Fig. 1 illustrates the whole process of this work, including data enrollment, automatic ROI definition, feature extraction, feature selection and model evaluation. The performance of tumor grading using mean APTw signals within four different ROIs was assessed (Fig. 2). As the histogram cutoff increased, the final ROI kept shrinking. The ROI with a histogram cutoff of 80% achieved the best performance, and thus was used in the following results. The ROC curves using top-ranked four radiomic features (red line) and only the conventional mean APTw signal (blue line) in a representative 5-fold cross-validation run are shown in Fig. 3a, illustrating the advantage of radiomics. This comparison was confirmed by 100 runs of 5-fold stratified cross-validation (Fig. 3b). In the pooled cross-validation results, the AUC and sensitivity of brain tumor grading achieved by the selected radiomic feature models were significantly higher than that from mean APTw signals (p=0.000), while the specificity did not exhibit significant statistical difference (p=0.652). Notably, the radiomic features also better explained (AUC=0.9134 vs. 0.8251) the training data than the mean APTw signal. Finally, the most frequent top-ranked features are shown in Fig. 5.

Discussion and Conclusion

Currently, most studies used mean APTw signals in various ROIs for grading in the APT MRI field. To the best of our knowledge, this is the first study that implemented a radiomic analysis of APTw images for grading brain tumors. The results demonstrated that the performance of the radiomic models was significantly (p=0.000) better than the conventional mean APTw signals. In conclusion, radiomics-based analysis of APTw images may facilitate the diagnosis and management of brain tumors.

Acknowledgements

NSFC grant number: 81971605, 61801421.