Journal: Scientific reports
This multi-center study developed a novel integrated model that combines deep learning-derived features and quantitative radiomics from multi-parametric MRI to enhance preoperative grading of gliomas.
The study analyzed 847 patients with histopathologically confirmed gliomas. The ensemble model incorporated:
- 3D CNN-extracted features
- Selected radiomic data
- Clinical variables
This integrated model outperformed both radiomics-only and deep learning-only approaches, achieving:
- An AUC of 0.946 for differentiating high-grade from low-grade gliomas
- Strong external validation with an AUC of 0.921
Additional findings include:
- High sensitivity and specificity in identifying IDH wild-type gliomas
- Accurate classification of nearly 90% of gliomas with aggressive molecular profiles but ambiguous imaging
This integrated, non-invasive framework offers improved preoperative risk stratification to guide surgical and treatment decisions. It also provides added interpretability by linking imaging biomarkers to tumor aggressiveness.