Machine learning decision tree models for multiclass classification of common malignant brain tumors using perfusion
Rodolphe Vallée1,2,3, Jean-Noël Vallée2,4, Carole Guillevin2,5
1Interdisciplinary Laboratory in Neurosciences, Physiology and Psychology (LINP2), Université Paris Lumière (UPL), Paris Nanterre University, Nanterre, France.
Frontiers in Oncology
|August 24, 2023
Summary
Machine learning decision trees accurately classify brain tumors using MRI perfusion and spectroscopy. This approach reveals key pathophysiological features, offering a hierarchical and personalized diagnostic tool for lymphomas, glioblastomas, and metastases.
Area of Science:
- Neuro-oncology
- Radiology
- Machine Learning
Background:
- Distinguishing between primary central nervous system lymphomas (PCNSLs), glioblastomas (GBMs), and brain metastases (METs) is critical for effective treatment.
- Traditional diagnostic methods can be invasive and time-consuming.
- Advanced imaging techniques offer non-invasive insights into tumor biology.
Purpose of the Study:
- To evaluate the efficacy of machine learning decision tree models in multiclass classification of PCNSLs, GBMs, and METs.
- To identify the key pathophysiological parameters derived from perfusion and spectroscopy MRI that drive the classification algorithm.
- To explore the potential for a hierarchical and personalized diagnostic approach using these models.
Main Methods:
- A cohort of 180 patients with histopathologically confirmed PCNSLs, GBMs, or METs underwent MRI between 2013 and 2020.
- Perfusion (rCBVmax, PSRmax) and spectroscopy (lac/Cr, Cho/NAA, Cho/Cr, lip/Cr) parameters were extracted.
- Classification and Regression Tree (CART) models were developed and validated using 5-fold cross-validation.
Main Results:
- The decision tree model achieved high classification performance: AUC of 0.98 for PCNSLs, 0.98 for GBMs, and 1.00 for METs, with an overall accuracy of 0.96.
- Key hierarchical features identified included Cho/NAA (proliferative/infiltrative characteristics), PSRmax (capillary permeability), and lac/Cr or Cho/Cr (metabolic activity/Warburg effect).
- Five distinct classification rules were extracted, demonstrating high predictive probability for tumor identification.
Conclusions:
- Machine learning decision tree models effectively classify brain tumors using perfusion and spectroscopy MRI data.
- The model's hierarchical structure highlights clinically relevant pathophysiological processes.
- This approach offers a promising, convenient, and personalized tool for brain tumor diagnosis.
Keywords:
classification and regression tree (CART)glioblastomalymphomametastasismulticlass classification
