Objectives: Given the glioblastoma (GBM) heterogeneity, survival-relevant high-risk subregions may exist and facilitate prognosis. The study aimed to identify the high-risk subregions on MRI, and to evaluate their survival stratification performance.
Methods: The gross tumor regions (GTRs) were delineated on the normalized MRI of 104 GBM patients. The signal intensity of voxels from 104 GTRs was pooled as global intensity vector, and K-means clustering was performed on it to find the optimal global clusters. Subregions were generated by assigning back voxels that belonged to each global cluster. Finally, a multiple instance learning (MIL) model was built and validated using radiomics features from each subregion. In this process, subregions predicted as positive would be treated as high-risk subregions, and patients with high-risk subregions inside the GTR would be predicted as having short-term survival.
Results: After K-means clustering, three global clusters were fixed and 294 subregions of 104 patients were generated. Then, the subregion-level MIL model was trained and tested by 200 (71 patients) and 94 subregions (33 patients). The accuracy, sensitivity, and specificity for survival stratification were 87.88%, 85.71%, and 89.47%. Furthermore, 41 high-risk subregions were correctly predicted from patients with short-term survival, in which the median overlap rate of non-enhancing component was 60%.
Conclusion: The stratification performance of high-risk subregions identified by the MIL model was higher than the GTR. The non-enhancing area on MRI was the most important component in high-risk subregions. The MIL approach provides a new perspective on the clinical challenges of glioma with coarse-grained labeling.
Key points: • The performance of high-risk subregions was more promising than the GTR for OS stratification. • The non-enhancing component was the most important in the high-risk subregions.
Keywords: Glioblastoma; Machine learning; Magnetic resonance imaging.