Adaptive Unsupervised Learning with Enhanced Feature Representation for Intra-tumor Partitioning and Survival Prediction for Glioblastoma

Abstract

Glioblastoma is profoundly heterogeneous in regional microstructure and vasculature. Characterizing the spatial heterogeneity of glioblastoma could lead to more precise treatment. With unsupervised learning techniques, glioblastoma MRI-derived radiomic features have been widely utilized for tumor sub-region segmentation and survival prediction. However, the reliability of algorithm outcomes is often challenged by both ambiguous intermediate process and instability introduced by the randomness of clustering algorithms, especially for data from heterogeneous patients.

In this paper, we propose an adaptive unsupervised learning approach for efficient MRI intra-tumor partitioning and glioblastoma survival prediction. A novel and problem-specific Feature-enhanced Auto-Encoder (FAE) is developed to enhance the representation of pairwise clinical modalities and therefore improve clustering stability of unsupervised learning algorithms such as K-means. Moreover, the entire process is modelled by the Bayesian optimization (BO) technique with a custom loss function that the hyper-parameters can be adaptively optimized in a reasonably few steps. The results demonstrate that the proposed approach can produce robust and clinically relevant MRI sub-regions and statistically significant survival predictions.

C. Li—Equal contribution.

5 Appendix

1.1 5.1 Details of Dataset and Imagine Processing

Patients with surgical resection (July 2010–August 2015) were consecutively recruited, with data prospectively collected by the multidisciplinary team (MDT) central review. All glioblastoma patients underwent pre-operative 3D MPRAGE (pre-contrast T1 and T1C), T2-weighted FLAIR, pMRI and dMRI sequences. All patients have a radiogical diagnosis of de novo glioblastoma, aged 18 to 75, eligible for craniotomy and radiotherapy, and all images resolution were resampled to $1\times 1 \times 1\,\mathrm{{m}}^3$.

Co-registration of the images was accomplished using the linear registration tool (FLIRT) included in the Oxford Centre for Functional MRI of the Brain Software Library (FSL) v5.0.0 (Oxford, UK) [5, 23]. NordicICE was used to process dynamic susceptibility contrast (DSC), one of the most frequently utilised perfusion methods (NordicNeuroLab). The arterial input function was automatically defined. The diffusion toolbox in FSL was used to process the diffusion images (DTI) [1]. The isotropic (p) and anisotropic (q) components were computed after normalisation and eddy current correction [20].

1.2 Details for Clinical Features

In this study, through the BO, the tumor were divided into 5 sub-regions as $\{\mathbf {P}_n\}_{n=1}^N$ from $\{\mathbf {Z}_{m^{\prime }}\}_{m^{\prime }=1}^M$, the features processed by the well-trained FAE, where $\mathbf {P}_n = \{\mathbf {p}_i\}_{i=1}^I$, $\mathbf {p}_i \in \{1,2,3,4,5\}$ denotes the sub-region labels for each pixel. Rather than representing the numerical grey value of images, the value of each $\mathbf {p}_i$ represents sub-region labels, rendering the majority of features in the GLCM and GLRLM families invalid. Finally, the Table 2 summarises the selected features which remain meaningful for the label matrix. Eventually, the clinical features $\{\mathbf {F}_n \}_{n=1}^N$, where $\mathbf {F}_n \in {\mathbb {R}}^{11 \times 1}$ include 9 spatial characteristics in Table 2 and the fraction of 2 significant sub-regions.

Table 2. Clinical features from GLCM matrix of size $N_g \times N_g$ and GLRLM matrix of size $N_g \times N_r$ family including Relative mutual information(RMI), Entropy, Joint Energy, Informational Measure of Correlation(IMC), Long Run Emphasis(LRE), Short Run Emphasis(SRE), Run Variance(RV) and Run Entropy(RE). $p(i,j|\theta )$ in the formula column describes the probability of the (i, j)th elements of matrices along angle $\theta$, $\mu =\sum _{i=1}^{N_g}\sum _{j=1}^{N_r}p(i,j|\theta )i$ denotes the average run length of GLRLM matrix [15].