(Sunday, 7/29/2018) 5:05 PM - 6:00 PM
Room: Room 205
Purpose: To accurately predict expression levels of Ki67 in glioma using machine learning models based on MR imaging.
Methods: Anatomic, diffusion, perfusion and permeability MRI were performed on 23 glioma patients and a total of 52 stereotactic biopsies were collected prior to resection as part of an IRB approved trial. Biopsy Ki67 expression levels were measured using immunohistochemical staining (% positive nuclei stained with MIB-1 antibody). Virtual biopsies in normal appearing white matter balanced the overrepresentation of tumor training data and were assumed to have 0% Ki67 expression.Images were registered to the T2 weighted image space and anatomic images were normalized to internal tissue reference standards. Imaging values were computed as the average intensity in a 5 mm diameter around the biopsy coordinates.We tested three imaging input selection methods (univariate testing, principal component analysis, and clinical intuition) and four common machine learning models (linear regression, neural network, decision tree, and random forest) using a repeated resampling scheme. The best model was chosen based on the lowest average test set error.
Results: Biopsies from all WHO grades (II-IV) were present in the final analysis. The random forest algorithm trained on four inputs (T2-weighted, fractional anisotropy, cerebral blood flow, and transfer constant kep) predicted Ki67 with an R-squared of 0.706 ± 0.185 and a root mean square error of 3.7%. The random forest was significantly better than the other models (paired t-test, p < 0.0001)
Conclusion: Ki67 can be predicted to clinically useful accuracies using clinically available MR imaging. This has implications for improving image guided treatment for glioma and justifies confirmatory imaging trials.
Keywords
Not Applicable / None Entered.
Taxonomy
Not Applicable / None Entered.
Contact Email
