Knowledge-guided gadolinium-free MRI radiomics predict 1p/19q co-deletion in IDH-mutant adult-type diffuse gliomas: A...

By integrating visual knowledge into radiomics, we aim to develop an interpretable and robust radiomic model based on Gadolinium-free MRI for preoperative identification of 1p/19q co-deletion in IDH-mutant adult-type diffuse gliomas.

MRI from 215 surgically confirmed IDH-mutant Adult-type Diffuse Glioma patients were collected from two centers and split into training, internal and external validation set. For knowledge-guided radiomics, features driven from T1WI, T2WI, and FLAIR demonstrating high interreader stability were extracted from the training set. Features significantly associated with four visual features (T2-FLAIR mismatch, homogeneity, enhancement, and calcification) were identified, and hierarchical clustering and Max-Relevance and Min-Redundancy were used to eliminate multicollinearity and redundancy. Stability selection using Lasso was performed for robust feature selection. Two classifiers (Support Vector Machines and Extreme Gradient Boosting) were trained with nested cross-validation for 1p/19q co-deletion prediction. For comparison, regular radiomic models without visual guidance were developed to evaluate the superiority of the knowledge-guided radiomics. This retrospective study was approved by the institutional review board of the centers and the requirement for informed consent was waived.

The knowledge-guided Support Vector Machines model achieved AUCs of 0.893 (95 %CI: 0.859-0.928) and 0.839 (95 % CI: 0.794-0.885) in the internal and external validation sets. The two regular radiomics models demonstrated various performance degradation, especially in the external set (AUCs: 0.681-0.767). The knowledge-guided radiomics model of either classifier outperformed corresponding regular radiomics across most performance metrics (p < 0.01) in the validation sets.

The knowledge-guided radiomic framework improves preoperative prediction of 1p/19q co-deletion in IDH-mutant adult-type diffuse gliomas, offering enhanced model interpretability and cross-center generalizability.