MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

MRI-Based Radiomics Distinguishes Treatment Effect From True Progression After Stereotactic Radiosurgery for Brain Metastases

L Peng1 , J Lee1 , K Sheikh1* , P Huang1 , V Parekh1 , B Baker1 , T Kirschbaum1 , F Silvestri1 , J Son1 , A Robinson1 , H Ames1 , J Grimm1 , L Chen1 , C Shen1 , M Soike2 , E McTyre2 , K Redmond1 , M Lim1 , M Jacobs1 , L Kleinberg1 , (1) Johns Hopkins University, Baltimore, MD, (2) Wake Forest School of Medicine, Winston Salem, NC

Presentations

(Sunday, 7/29/2018) 5:05 PM - 6:00 PM

Room: Karl Dean Ballroom C

Purpose: Growth due to treatment effect after stereotactic radiosurgery (SRS) for brain metastases (BM) is a common phenomenon often indistinguishable from true progression (TP) on imaging. Radiomics is a reemerging field that promises to improve upon conventional imaging. In this study, we hypothesized that a radiomics-based prediction model would outperform a prediction model based solely on clinical data in distinguishing true progression after SRS.

Methods: Patients at a single institution treated with SRS for BM from 2003-2017 were reviewed to identify cases with subsequent resection for suspected progression. Additional cases of presumed treatment effect were also included where lesions grew but subsequently regressed spontaneously. Radiomic features of contoured lesions on T1 post contrast and FLAIR MRI sequences were extracted. A feature subset was then selected based on minimizing inter-correlations and utilized to train random forest (RF) classifiers. Performance of this radiomics-based classifier (RadiomicRF) was assessed with leave-one-out cross-validation (LOOCV) and compared with the LOOCV performance of a clinical RF classifier (ClinicalRF) trained on clinical features.

Results: We identified 82 treated lesions across 66 patients, with 77 lesions having pathologic confirmation. At least some component of TP was identified in 52/82 (63%) cases. There were 1904 radiomic features extracted per contoured lesion. Ten features with minimal redundancy and low correlation were ultimately selected for training RadiomicRF. On LOOCV, ClinicalRF had an accuracy of just 58% with area under the receiver operating characteristic curve (AUC) of 0.59. By contrast, RadiomicRF on LOOCV had 77% accuracy and AUC 0.84. The radiomic feature most predictive for TP was gray level size-zone matrix (GLSZM) non-uniformity.

Conclusion: Radiomics showed great promise for differentiating between treatment effect and TP in brain lesions treated with SRS and outperformed a model based on clinical data alone. These results warrant further validation in independent datasets.

Keywords

Brain, MRI, Feature Extraction

Taxonomy

TH- response assessment : Machine learning

Contact Email