Unsupervised machine learning using K-means identifies radiomic subgroups of pediatric low-grade gliomas that correlate with key molecular markers

Published in
Debanjan Haldar, Anahita Fathi Kazerooni, Sherjeel Arif, Ariana Familiar, Rachel Madhogarhia, Nastaran Khalili, Sina Bagheri, Hannah Anderson, Ibraheem Salman Shaikh, Aria Mahtabfar, Meen Chul Kim, Wenxin Tu, Jefferey Ware, Arastoo Vossough, Christos Davatzikos, Phillip B. Storm, Adam Resnick, Ali Nabavizadeh


This study aimed to investigate if an unsupervised machine learning approach based on radiomic features could reveal distinct imaging subtypes of pediatric low-grade gliomas (pLGGs). The researchers collected multi-parametric MR images from 157 patients with pLGGs and extracted quantitative radiomic features from the tumorous region. They then used K-means clustering to identify three distinct imaging-based subtypes. The subtypes differed in mutational frequencies of BRAF as well as the gene expression of BRAF. The study also found significant differences in age, tumor location, and tumor histology between the imaging subtypes. This study suggests that clustering of pLGGs based on radiomic features could enhance our ability to better characterize pLGGs.