MRI-based whole-tumor radiomics to classify the types of pediatric posterior fossa brain tumor

Machine Learning in the Classification of Pediatric Posterior Fossa Tumors: A Systematic Review

“…Table 1 features study method data from all 25 studies of the MLAs applied to the classification of pediatric PFTs. Twenty-two papers used imaging data to classify PFTs, including both non-contrast and contrast-enhanced T1/T2-weighted MRI, DWI, and MR-spectroscopy [ 16 , 17 , 18 , 19 , 20 , 21 , 22 , 23 , 24 , 25 , 26 , 27 , 28 , 29 , 30 , 31 , 32 , 33 , 34 , 35 , 36 , 37 ]. Three papers used molecular methods to classify these tumors based on microscopy slides or methylation array data [ 38 , 39 , 40 ].…”

“…Study populations varied significantly, ranging from cohorts of 23 patients to 617 patients [ 31 , 40 ]. Pilocytic astrocytoma and medulloblastoma were the most well-represented PFTs across all reports with inclusion in 19 and 22 studies, respectively [ 17 , 18 , 19 , 20 , 21 , 22 , 23 , 24 , 25 , 26 , 27 , 28 , 29 , 30 , 31 , 32 , 34 , 35 , 36 , 37 , 38 , 39 , 40 ]. Ependymomas, while included in most studies, had a small individual sample size per study, with many analyses including fewer than 20 ependymoma patients in training or validation datasets [ 16 , 17 , 18 , 21 , 22 , 23 , 27 , 29 , 30 , 32 , 36 ].…”

“…Pilocytic astrocytoma and medulloblastoma were the most well-represented PFTs across all reports with inclusion in 19 and 22 studies, respectively [ 17 , 18 , 19 , 20 , 21 , 22 , 23 , 24 , 25 , 26 , 27 , 28 , 29 , 30 , 31 , 32 , 34 , 35 , 36 , 37 , 38 , 39 , 40 ]. Ependymomas, while included in most studies, had a small individual sample size per study, with many analyses including fewer than 20 ependymoma patients in training or validation datasets [ 16 , 17 , 18 , 21 , 22 , 23 , 27 , 29 , 30 , 32 , 36 ]. Less common pathologies, such as embryonal tumors, gangliogliomas, atypical teratoid rhabdoid tumors, and others, were heterogeneously studied and only featured in a small minority of reports [ 26 , 28 , 38 ].…”

“…Imaging data were segmented to identify regions of interest using a combination of manual and semiautomatic methods. Eighteen of the papers then included a normalization step in which imaging data were standardized to minimize noise [ 17 , 18 , 19 , 20 , 21 , 22 , 23 , 25 , 26 , 27 , 28 , 29 , 31 , 33 , 36 , 37 , 39 , 40 ]. Features were extracted from the relevant imaging modalities with some papers yielding as few as 13 features while others generated over 11,000 [ 20 , 39 ].…”

“…Features were extracted from the relevant imaging modalities with some papers yielding as few as 13 features while others generated over 11,000 [ 20 , 39 ]. Nineteen studies employed various methods of feature selection to decrease feature dimensionality for the final analysis [ 19 , 20 , 21 , 22 , 23 , 24 , 25 , 26 , 27 , 29 , 30 , 32 , 33 , 34 , 35 , 36 , 37 , 38 , 40 ]. A machine learning classifier was then applied to discriminate between tumor types based on extracted features.…”