Radiomic Based Machine Learning Performance for a Three Class Problem in Neuro-Oncology: Time to Test the Waters?

Area of Science:

• Radiology and Imaging Science
• Oncology
• Medical Artificial Intelligence

Background:

• Previous radiomics studies primarily focused on two-class brain tumor classification, limiting their real-world applicability.
• Differentiating between glioblastoma (GBM), primary central nervous system lymphoma (PCNSL), and metastatic disease using radiomics requires further investigation.
• The impact of MRI sequence selection and tumor subregion segmentation on radiomics model performance is not well-understood.

Purpose of the Study:

• To assess the performance of radiomics in distinguishing between GBM, PCNSL, and metastatic brain tumors.
• To evaluate factors influencing radiomics model performance, including MRI sequence choice and segmentation strategies.
• To compare the utility of single MRI sequences versus multiparametric MRI (MP-MRI) for brain tumor classification.

Main Methods:

• A retrospective analysis of 253 patients with GBM, PCNSL, or metastatic brain tumors.
• Radiomic features were extracted using two pipelines: whole tumor (enhancing + necrotic) and edema masks, and separate enhancing, necrotic, and edema masks.
• Model performance was evaluated using MP-MRI, individual sequences, and the T1 contrast-enhanced (T1-CE) sequence across various model/feature selection combinations.

Main Results:

• The second radiomics pipeline, segmenting distinct tumor subregions, demonstrated high performance (Brier score: 0.311-0.325).
• The Gradient Boosting Regression Model (GBRM) using the full feature set from the T1-CE sequence achieved the best performance.
• Top-performing models showed no significant difference between MP-MRI (AUC 0.910) and T1-CE sequences with (AUC 0.908) or without edema masks (AUC 0.894), indicating T1-CE's standalone efficacy.

Conclusions:

• The T1 contrast-enhanced (T1-CE) MRI sequence alone provides performance comparable to multiparametric MRI (MP-MRI) for differentiating the three most common malignant brain tumors.
• Radiomics models utilizing detailed tumor subregion segmentation (enhancing, necrotic, edema) show superior performance.
• Model performance is influenced by tumor subregion definition and the choice of modeling and feature selection techniques.