Abstract
Purpose
We used radiomics-based machine learning (ML) of T2-weighted imaging (T2WI) and contrast-enhanced T1-weighted imaging (T1C) for assessing relapse risk in patients with high-grade meningiomas (HGMs).
Methods
279 features were extracted from each ROI. The datasets were randomly divided into two groups, the training set (∼70%) and the test set (∼30%). Data of 192 individuals were used for external validation. Combinations of data preprocessing methods, including normalization (Min-Max, Z-score, Mean) and dimensionality reduction (Pearson Correlation Coefficients (PCC)), and feature selector (max-Number, cluster) were analyzed for their prediction performance (totaling to 60 combinations). Kaplan–Meier curve, Cox proportional hazards regression model were used and concordance index (C-index), integrated Brier score (IBS) were selected.
Results
WHO grade, age, gender, histogram (Mean, Perc.90%, Perc.99%), Gray-level co-occurrence matrix (S(3, -3)DifVarnc, S(5, 5)Correlat, S(1, 0)SumEntrp, S(2, -2)InvDfMom), Teta1, WavEnLL_s-2 and GrVariance were identified as the significant recurrence factors. The pipeline using Mean_PCC_Cluster_10 of T1C yielded the highest efficiency with an IBS of 0.170, 0.188, 0.208 and C-index of 0.709, 0.705, 0.602 in the train, test and validation sets, respectively. The pipeline using MinMax_PCC_Cluster_19 of T2WI yielded the highest efficiency with an IBS of 0.189, 0.175, 0.185 and C-index of 0.783, 0.66, 0.649 in the train, test and validation sets. The pipeline using MinMax_PCC_Cluster_13 of T2WI + T1C yielded the highest efficiency with an IBS of 0.152, 0.164, 0.191 and C-index of 0.701, 0.656, 0.593 in the train, test and validation sets, respectively.
Conclusion
Machine learning on MRI radiomic features can slightly help predict recurrence risk in HGMs. T2WI or T1C yielded better efficiency than T2WI + T1C. The parameters with the best power were Mean, Perc.99%, WavEnLL_s-2, Teta1 and GrVariance.