Deep Learning and Habitat Radiomics for the Prediction of Glioma Pathology Using Multiparametric MRI: A Multicenter...

Area of Science:

• Neuro-oncology
• Medical Imaging
• Artificial Intelligence

Background:

• Glioma pathological outcome prediction is crucial for treatment planning.
• Tumor heterogeneity limits the predictive accuracy of current radiomics studies.
• Novel methods are needed to improve prediction of glioma characteristics.

Purpose of the Study:

• To enhance pathological prediction outcomes in glioma by integrating habitat analysis with deep learning.
• To identify optimal models for predicting glioma grades, Ki67 expression, P53 mutation, and IDH1 mutation.

Main Methods:

• Collected MR imaging (T1 contrast-enhanced, T2-weighted) and pathological data from 387 primary glioma cases across three hospitals.
• Employed radiomics, deep learning (DenseNet161, ResNet50, Inception_v3), and habitat analysis techniques.
• Developed and compared various models including LightGBM, SVM, and MLP, integrating imaging and clinical features.

Main Results:

• Habitat+Deep Learning models achieved optimal prediction for glioma grades and Ki67 levels.
• Deep Learning models were optimal for P53 mutation prediction.
• A combination of Habitat+Radiomics models excelled in predicting IDH1 mutation.

Conclusions:

• The integration of habitat analysis and deep learning significantly improves the prediction of key glioma pathological features.
• Different modeling approaches show optimal performance for distinct predictive tasks, highlighting the complexity of glioma biology.
• These findings suggest a promising avenue for more accurate and personalized glioma management.