| Abstrakt: |
Texture-based convolutional neural networks (CNNs) have shown great promise in predicting various types of cancer, including lower grade glioma (LGG) through radiomics analysis. However, the use of CNN-based radiomics requires a large training set to avoid overfitting. To overcome this problem, the study proposes a novel panel of radiomic/texture features based on principal component analysis (PCA) applied to pretrained CNN features. The study used extracted PCA-CNN radiomic features from multimodal magnetic resonance imaging (MRI) images as input to a random forest (RF) classifier to predict immune cell markers, the gene status, and the survival outcome for LGG patients (n = 83). The results of the experiments demonstrate that RF with PCA-CNN radiomic features improved the classification performance, achieving the highest significant classification between short- and long-term survival outcomes. Notably, the area under the curve for PCA-CNN radiomic features with RF was 78.53% (p = 0.0008), which was significantly better than using gene status 63.14% (p = 0.23), clinical variables 52.60% (p = 0.32), standard radiomic features 72.56% (p = 0.02), immune cell markers 65.67% (p = 0.007), conditional entropy 74.54% (p = 0.0058), Gaussian mixture model-CNN 74.94% (p = 0.0053), or using 3D CNN classification directly without RF 72.61% (p = 0.01). The proposed PCA-CNN-based radiomic model outperformed state-of-the-art techniques to predict the survival outcome of LGG patients. [ABSTRACT FROM AUTHOR] |
Full text from SpringerLink