ABSTRACT
This study aimed to explore the value of deep learning (DL)-assisted quantitative susceptibility mapping (QSM) in glioma grading and molecular subtyping. Forty-two patients with gliomas, who underwent preoperative T2 fluid-attenuated inversion recovery (T2 FLAIR), contrast-enhanced T1-weighted imaging (T1WI + C), and QSM scanning at 3.0T magnetic resonance imaging (MRI) were included in this study. Histopathology and immunohistochemistry staining were used to determine glioma grades, and isocitrate dehydrogenase (IDH) 1 and alpha thalassemia/mental retardation syndrome X-linked gene (ATRX) subtypes. Tumor segmentation was performed manually using Insight Toolkit-SNAP program (www.itksnap.org). An inception convolutional neural network (CNN) with a subsequent linear layer was employed as the training encoder to capture multi-scale features from MRI slices. Fivefold cross-validation was utilized as the training strategy (seven samples for each fold), and the ratio of sample size of the training, validation, and test dataset was 4:1:1. The performance was evaluated by the accuracy and area under the curve (AUC). With the inception CNN, single modal of QSM showed better performance in differentiating glioblastomas (GBM) and other grade gliomas (OGG, grade II-III), and predicting IDH1 mutation and ATRX loss (accuracy: 0.80, 0.77, 0.60) than either T2 FLAIR (0.69, 0.57, 0.54) or T1WI + C (0.74, 0.57, 0.46). When combining three modalities, compared with any single modality, the best AUC/accuracy/F1-scores were reached in grading gliomas (OGG and GBM: 0.91/0.89/0.87, low-grade and high-grade gliomas: 0.83/0.86/0.81), predicting IDH1 mutation (0.88/0.89/0.85), and predicting ATRX loss (0.78/0.71/0.67). As a supplement to conventional MRI, DL-assisted QSM is a promising molecular imaging method to evaluate glioma grades, IDH1 mutation, and ATRX loss. Supplementary Information: The online version contains supplementary material available at 10.1007/s43657-022-00087-6.
ABSTRACT
OBJECTIVE: To predict silent corticotroph adenomas (SCAs) among non-functioning pituitary adenomas preoperatively using noninvasive radiomics. METHODS: A total of 302 patients including 146 patients diagnosed with SCAs and 156 patients with non-SCAs were enrolled (training set: n = 242; test set: n = 60). Tumor segmentation was manually generated using ITK-SNAP. From T1-weighted imaging (T1WI), T2-weighted imaging (T2WI), and contrast-enhanced T1WI, 2550 radiomics features were extracted using Pyradiomics. Pearson's correlation coefficient values were calculated to exclude redundant features. Several machine learning algorithms were developed to predict SCAs incorporating the radiomics and semantic features including clinical, laboratory, and radiology-associated features. The performance of models was evaluated by AUC. RESULTS: Patients in the SCA group were younger (49.5 vs 55.2 years old) and more female (85.6% vs 37.2%) than those in the non-SCA group (p < 0.001). More invasiveness (p = 0.011) and cystic and microcystic change (p < 0.001) were observed in patients with SCAs. The ensemble algorithm presented the largest AUC of 0.927 among all the algorithms trained in the test set, and the accuracy, specificity, and sensitivity of predicting SCAs were all 0.867 (at cut-off 0.5). The overall model performed better than that only using semantic features available in the clinic. Radiomics prediction was the most important feature, with gender ranking second and age ranking third. Radiomics features on T2WI were superior to those on other MR modalities in SCA prediction. CONCLUSION: Our ensemble learning model outperformed current clinical practice in differentiating patients with SCAs and non-SCAs using radiomics, which might help make appropriate treatment strategies. KEY POINTS: ⢠Radiomics might improve the preoperative diagnosis of SCAs by MR images. ⢠T2WI was superior to T1WI and CE-T1WI in the preoperative diagnosis of SCAs. ⢠The ensemble machine learning model outperformed current clinical practice in SCAs diagnosis and treatment decision-making could be more individualised using the nomogram.
