Research Review on the Predictive Value of Low-Grade Glioma Postoperative Growth Models

Introduction

Glioma is an aggressive brain tumor whose cells rapidly diffuse within the brain. Understanding and predicting the pattern and speed of this diffusion can help optimize treatment plans. Glioma growth models based on diffusion-proliferation have shown feasibility, but applying and evaluating these models in actual clinical data remain challenging. To improve this evaluation, this study proposes treating the tumor growth problem as a ranking issue and using Average Precision (AP) as a metric. This method does not require specific volume thresholds and can more accurately assess spatial patterns.

Study Source

This paper is authored by Karin A. van Garderen, Sebastian R. van der Voort, Maarten M. J. Wijnenga, and others from the departments of Radiology and Nuclear Medicine, Neurosurgery, Pathology, and Neurology of Erasmus Medical Center in Rotterdam, the Netherlands. The paper is published in the January 2024 issue (Vol. 43, No. 1) of “IEEE Transactions on Medical Imaging” and is partially funded by the GLASS-NL project of the Dutch Cancer Society, the Delta Med project, the Netherlands Research Council (NWO), and the European Union’s Horizon 2020 Research and Innovation Program.

Research Methods

Tumor Growth Model

The study uses tumor growth models based on the diffusion-proliferation model, including homogeneous isotropic diffusion models and models that consider anisotropic diffusion components. The partial differential equation defined is as follows:

$$\frac{dc}{dt} = \nabla(D\nabla c) + \rho c(1 - c)$$

where $\rho$ is the growth factor, and $D$ is the diffusion tensor. The model considers the boundaries between the brain and cerebrospinal fluid, formulated as:

$$D = \kappa(x)I + \tau f(x)T(x)$$

where $\kappa$ and $\tau$ are different weighting parameters, $f(x)$ is the local fractional anisotropy, and $T$ is the normalized diffusion tensor obtained from the diffusion tensor imaging.

Dataset and Image Processing

The model is applied to 14 patients with low-grade glioma (LGG) who did not receive other treatments post-surgery. Image processing steps include registering patient images to a healthy brain template and manually segmenting tumor boundaries. During preprocessing, the elastix software was used for image registration, and nonrigid deformation was employed to capture brain deformation caused by tumor growth and surgery.

Model Selection and Evaluation Metrics

Three models were designed in this study: a basic model, a tissue segmentation-based model, and a DTI-based model. Model parameters were chosen in the healthy brain template to ensure that each model generates significantly different tumor shapes. The study proposes treating tumor growth prediction as a spatial invasion ranking problem, using AP as the evaluation metric. The AP metric is independent of the timeline, separating spatial accuracy from the timeline.

Research Results

Simulation Experiments

Simulation experiments compared the tumor morphology and growth speed in three different initial positions, finding similar growth speeds and effective diffusion across the models. Model morphology showed that the DTI-based model exhibited higher accuracy in predictions, especially in long-distance invasion predictions.

Patient Data

In patient data, the DTI-based anisotropic diffusion model significantly outperformed the basic model and the tissue segmentation-based model, particularly in predicting recurrence tumor shapes closer to actual conditions. When considering initial tumors, the models showed no significant difference, but the DTI-based model still outperformed in some aspects.

Conclusion and Significance

This study proposes a new evaluation framework for the tumor growth problem and uses the AP metric for assessment. In the prediction of postoperative growth of low-grade glioma, the DTI-based anisotropic diffusion model significantly improved the accuracy of recurrent tumor shape prediction. Through this new method and the public release of code and data, the study provides a better basis for comparison for future improvements in glioma growth models. This work demonstrates not only the advantages of the AP metric in evaluating glioma growth models but also emphasizes the importance of data transparency and standardization in model development.

Research Highlights and Future Prospects

1. Innovative Evaluation Method: The study is the first to treat the tumor growth prediction problem as a ranking problem and use AP as an evaluation criterion, helping to more accurately assess spatial growth patterns.

2. Model Improvement: Results show that the DTI-based model performs excellently in predicting recurrent tumors, providing a theoretical basis for developing more complex growth models in the future.

3. Data Sharing: By making code and data public, the study promotes future innovation and standardization in glioma growth model research.

This study introduces a new method for evaluating the growth models of low-grade gliomas, making predictions more accurate, and holds significant potential and value for clinical applications.