One class classification-empowered radiomics for noninvasively accurate prediction of glioma isocitrate dehydrogenase mutation using multiparametric magnetic resonance imaging

Abstract

Background

Noninvasive detection of isocitrate dehydrogenase (IDH) mutations is crucial for preoperative decision-making in patients with glioma. While radiomics has been applied, data imbalance—specifically between IDH wild-type and mutated genes—remains underexplored. We developed a one-class classification-empowered radiomics (OCCR) model, trained exclusively on IDH wild-type patients, to distinguish them from IDH mutation cases.

Materials and Methods

This study included 495 patients from the UCSF Preoperative Diffuse Glioma MRI dataset. T1, T1ce, and FLAIR sequences were registered to T2 and resampled to a 1-mm isotropic resolution. The coregistered data were skull-stripped, and the tumor region was segmented using an ensemble model, followed by manual refinement. We extracted 386 radiomics features from the four MRI sequences and input them into an auto-encoder with 7 hidden layers for reconstruction. The OCCR model was trained on wild-type IDH patients, using the mean square error between the original and reconstructed features as guidance. During validation, reconstruction error was used to differentiate IDH mutations from the wild type.

Results

The hold-out validation demonstrated that OCCR performance improved as the number of training samples increased, achieving a peak area under the receiver operating characteristic curve of 0.8018. Visualization of reconstruction errors highlighted first-order and gray-level co-occurrence matrix features in the T1ce sequence.

Conclusions

This study demonstrates the feasibility of integrating one-class classification into radiomics for the determination of preoperative IDH mutation status in patients with glioma using multiparametric MRI. This versatile model holds potential for other diseases with substantial data imbalance.