MENU

Click here to

×

Are you sure ?

Yes, do it No, cancel

MRI Radiomic Analysis for Survival Prediction in Diffuse Midline Glioma

L Tam1, M Han1, D Yecies1, K Yeom1, S Mattonen2*, (1) Stanford University, Stanford, CA, USA (2) Western University, London, ON, CA

Presentations

(Sunday, 7/12/2020)   [Eastern Time (GMT-4)]

Room: AAPM ePoster Library

Purpose: Diffuse midline gliomas (DMG), formerly diffuse intrinsic pontine glioma (DIPG), are lethal pediatric brain tumors with dismal prognoses. Presently, magnetic resonance imaging (MRI) is the mainstay of disease diagnosis and surveillance. We aimed to identify prognostic imaging based radiomic biomarkers of DMG and compare their performance to standard clinical variables at presentation.


Methods: We retrospectively analyzed treatment naïve children with DMG from five centers. We manually isolated tumor volumes on T1-post-contrast (T1-post) and T2-weighted (T2) diagnostic MRIs. Pyradiomics was used to extract high-dimensional features on the original, wavelet, and Laplacian of Gaussian (LoG) filtered images. A total of 900 features were extracted on each image series, including first-order statistics, size, shape, and texture-based features. Overall survival (OS) served as our outcome of interest. 10-fold cross-validation of least absolute shrinkage and selection operator (LASSO) Cox regression was used to predict OS. The performance of each Cox model was assessed using the concordance (C) index. We compared model performance using clinical variables only (age at diagnosis and sex), radiomics only, and radiomics plus clinical variables.


Results: A total of 104 children were analyzed (46 males; median age = 6.5 years; median OS = 11 months). Nine radiomic features were selected from both T1-post and T2 imaging. Two features were from T1-post (2 texture [wavelet]), whereas the remaining seven were from T2 (5 first-order [1 original and 4 wavelet] and 2 texture [1 wavelet and 1 LoG]). This model demonstrated significantly higher performance than the model with only clinical features (C: 0.68 vs. 0.59, p<0.001). Furthermore, adding clinical features to the radiomic features slightly improved prediction, but this was not significant (C: 0.70 vs. 0.68, p=0.06).


Conclusion: Our pilot study shows a potential role for MRI-based radiomics and machine learning for DMG risk stratification and as imaging-based biomarkers for clinical therapy trials.

Keywords

MRI, Quantitative Imaging, Brain

Taxonomy

IM/TH- Image Analysis (Single Modality or Multi-Modality): Computer-aided decision support systems (detection, diagnosis, risk prediction, staging, treatment response assessment/monitoring, prognosis prediction)

Contact Email