Accurate Patient-Specific Machine Learning Models of Glioblastoma Invasion Using Transfer Learning.

Area of Science:

• Radiology
• Oncology
• Medical Imaging Analysis

Background:

• Glioblastoma treatment relies on accurate tumor cell density modeling.
• Interpatient variability in glioblastoma poses challenges for predictive modeling.
• Current MR imaging models struggle with individual patient differences.

Purpose of the Study:

• To develop an individualized transfer learning method for glioblastoma tumor cell density prediction.
• To improve the accuracy of MR imaging-based models by accounting for interpatient variability.
• To enhance targeted glioblastoma treatment strategies.

Main Methods:

• Recruited glioblastoma patients for MR imaging (CE-MR, DSC-MR, DTI) and biopsies.
• Assessed tumor cell density and correlated with MR imaging parameters (e.g., relative CBV, MD, FA).
• Developed and compared generalized 'one-model-fits-all' and individualized transfer learning predictive models.

Main Results:

• Tumor cell density correlated with relative CBV (r=0.33) and postcontrast T1 (r=0.36).
• Transfer learning significantly improved univariate (r=0.53) and multivariate (r=0.88) predictive performance over one-model-fits-all approaches.
• Multivariate transfer learning achieved a mean absolute error of 5.66%.

Conclusions:

• Individualized transfer learning models significantly enhance glioblastoma tumor cell density prediction.
• This method effectively addresses interpatient variability in MR imaging-based modeling.
• Improved predictive accuracy supports better-targeted glioblastoma therapies.