Objective:

To evaluate the effectiveness of AI models in noninvasive identification of pathogenic variants in CNS tumors using genetic and imaging data.

Approach:

• Model Development: Two machine-learning models were developed: a dense neural network for mutation data from 12 key genes using next-generation sequencing of CSF-derived circulating tumor DNA, and a convolutional neural network for standardized MRI images.
• Data Sources: Data were collected from paired tissue, CSF, and MRI datasets from patients with CNS tumors.
• Model Combination: Averaging or majority voting was used to combine the outputs of the models to optimize diagnostic robustness.

Key Findings:

• The models classified samples into three categories: glial tumors (GBM), nonglial CNS tumors, and normal tissue with strong reliability.
• Misclassifications primarily occurred between glial and nonglial tumors, while normal samples were consistently identified.
• The dense neural network achieved a Matthews Correlation Coefficient (MCC) of 0.8822, and the MRI-based convolutional network achieved an MCC of 0.8525.
• Combining the outputs of both models improved performance in identifying pathogenic variants and tumor types.

Interpretation:

Conclusion:

Sources:

• Association for Molecular Pathology