Abstract
Glioblastoma exhibits significant intratumoral heterogeneity, making precise treatment planning difficult. This study proposes an adaptive unsupervised learning framework for robust tumor sub-region partitioning using multi-modal MRI. The framework integrates a Feature-enhanced Auto-Encoder (FAE) to capture pairwise modality correlations and a Bayesian Optimization (BO) loop to optimize hyper-parameters for both clustering stability and clinical relevance. Results on 117 patients demonstrate superior stability (0.923) and significant survival prediction ($p < 0.005$ in the test set), offering a robust tool for clinical decision support.
Introduction
Glioblastoma (GBM) is a highly aggressive brain tumor characterized by extreme internal heterogeneity. Different regions within the same tumor can exhibit varying microstructures, blood supply, and metabolic activity. Identifying these "tumor habitats" via multi-modal MRI is critical for targeted therapy and prognosis. However, traditional unsupervised methods like K-means are often unstable, sensitive to noise, and require manual selection of cluster numbers, leading to inconsistent results across patient cohorts. This research addresses this gap by jointly optimizing for clustering stability and clinical relevance.
Methodology
The proposed framework introduces two primary innovations: the Feature-enhanced Auto-Encoder (FAE) and an adaptive Bayesian Optimization (BO) loop.
Feature-enhanced Auto-Encoder (FAE)
Unlike standard auto-encoders that compress all modalities simultaneously, the FAE utilizes a pairwise architecture. For $M$ MRI modalities, it constructs encoders for every possible pair, forcing the network to learn shared features between specific clinical parameters (e.g., perfusion and diffusion). A dual decoding strategy is employed: individual decoders reconstruct pairwise information, while a global decoder ($D_s$) ensures the latent features maintain overall correlation. This approach effectively filters noise and highlights clinically grounded features.
Adaptive Bayesian Optimization
The pipeline is treated as a "black box" where hyper-parameters—specifically the outlier threshold ($\gamma$) and the number of clusters ($\eta$)—are optimized using BO. The optimization is guided by a joint loss function: $$\mathcal{L} = \alpha L_s + (1-\alpha) L_p$$ where $L_s$ represents stability loss (measured via Hamming distance across data splits) and $L_p$ represents significance loss (based on the p-value of survival predictions). The iterative process for finding optimal hyper-parameters is detailed in Algorithm 1.
Experimental Setup
The study utilized multi-modal MRI data from 117 glioblastoma patients, partitioned into a training set (82 patients) and a test set (35 patients). The modalities included relative cerebral blood volume (rCBV) and the isotropic ($p$) and anisotropic ($q$) components of diffusion MRI.
Results and Discussion
The FAE-based framework demonstrated superior performance across multiple metrics:
- Stability: The FAE achieved a stability score of 0.923 for 4 clusters, outperforming the baseline auto-encoder (0.890).
- Cluster Quality: The Calinski-Harabasz (CH) score for the FAE was nearly three times higher than the baseline, indicating significantly better cluster separation.
- Clinical Significance: The optimized model successfully stratified patients into high-risk and low-risk groups. In the training set, the survival separation reached a p-value of 0.013, which improved to 0.0034 in the independent test set.
- Efficiency: The BO framework typically converged within 6 to 10 iterations, making it a computationally efficient solution for hyper-parameter tuning in clinical settings.
Conclusion
This study provides a robust, automated framework for glioblastoma sub-region partitioning. By integrating stability and clinical significance into the optimization loop, the method overcomes the traditional limitations of unsupervised clustering. Future work will explore the inclusion of additional MRI modalities and non-imaging clinical variables, such as genetic markers, to further enhance predictive accuracy.
Summary updated