Radiogenomic classification for MGMT promoter methylation status using multi-omics fused feature space for least invasive diagnosis through mpMRI scans

Abstract

Accurate radiogenomic classification of brain tumors is important to improve the standard of diagnosis, prognosis, and treatment planning for patients with glioblastoma. In this study, we propose a novel two-stage MGMT Promoter Methylation Prediction (MGMT-PMP) system that extracts latent features fused with radiomic features predicting the genetic subtype of glioblastoma. A novel fine-tuned deep learning architecture, namely Deep Learning Radiomic Feature Extraction (DLRFE) module, is proposed for latent feature extraction that fuses the quantitative knowledge to the spatial distribution and the size of tumorous structure through radiomic features: (GLCM, HOG, and LBP). The application of the novice rejection algorithm has been found significantly effective in selecting and isolating the negative training instances out of the original dataset. The fused feature vectors are then used for training and testing by k-NN and SVM classifiers. The 2021 RSNA Brain Tumor challenge dataset (BraTS-2021) consists of four structural mpMRIs, viz. fluid-attenuated inversion-recovery, T1-weighted, T1-weighted contrast enhancement, and T2-weighted. We evaluated the classification performance, for the very first time in published form, in terms of measures like accuracy, F₁-score, and Matthews correlation coefficient. The Jackknife tenfold cross-validation was used for training and testing BraTS-2021 dataset validation. The highest classification performance is (96.84 ± 0.09)%, (96.08 ± 0.10)%, and (97.44 ± 0.14)% as accuracy, sensitivity, and specificity respectively to detect MGMT methylation status for patients suffering from glioblastoma. Deep learning feature extraction with radiogenomic features, fusing imaging phenotypes and molecular structure, using rejection algorithm has been found to perform outclass capable of detecting MGMT methylation status of glioblastoma patients. The approach relates the genomic variation with radiomic features forming a bridge between two areas of research that may prove useful for clinical treatment planning leading to better outcomes.

Figure 1

The proposed MGMT promoter methylation prediction (MGMT-PMP) system based on DLRFE forming a two-stage HFS.

Figure 2

BraTS-2021 mpMRI scans from top to bottom: sagittal, axial and coronal view: first, second and third columns show (a,e,i) T1w, (b,f,j) T1wCE, (c,g,k) FLAIR, and (d,h,l) T2w images respectively.

Figure 3

Efficient preprocessing steps for BraTS-2021 for mpMRI scans to MGMT methylation prediction stage.

Figure 4

Pseudocode for the file rejection algorithm (RA).

Figure 5

Latent feature extraction $x_{\mathit{latent}}$ using DLRFE-module via feature-bleeding through fully connected layers.

Figure 6

A simple diagram for GLCM computation; (a) input matrix, (b) GLCM computed from (a) using d = 1 at 0° orientation, and (c) set of orientation, θ.

Figure 7

Schematic diagram of FEM for HOG features; (a) a rescaled mpMRI scan for radiogenomic dataset, (b) schematic HOG cells and blocks superimposed on (a), (c) HOG descriptors, (d) cell and block representation, and (e) the magnified view of a single cell.

Figure 8

Schematic diagram of LBP operation to find the discriminative features.

Figure 9

Confusion Matrices (based on the best of ten) for the k-NN classifier (k = 1) using proposed DLRFE module for backbone learning using BraTS-2021 dataset with RA to improve the discrimination of the features using epochs as: (a) 1, (b) 5, (c) 10, (d) 25, (e) 30, and (f) 50.

Figure 10

Performance analysis of RA for BraTS-2021 dataset based on: (a) accuracy variation with DL with softmax used as the classifier, (b) accuracy variation using k-NN model ML with x-axis showing the epochs used to retrieve the $x_{\mathit{latent}}$ (c) F₁-score plot, (d) MCC plot, (e) AUC (ROC) variation with epochs, (f) precision variation with epochs, (g) recall variation with epochs, (h) PPV and NPV plots for varying epochs without RA, (i) NPV with and without RA plots.

Figure 11

The scatter plot for latent features mapped to ℜ² using the t-SNE technique illustrating Stage 1 for classification of mpMRI scans into MGMT+ (“1”) and MGMT− (“0”) classes for the original dataset: (a) FC1 features $x_{FC1}$, (b) FC2 features $x_{FC2}$, and (c) latent features $x_{\mathit{latent}}$. The results depicting visual classification with the rejection algorithm are: (d) $x_{FC1}^{{}^{\prime }}$ (e) $x_{FC2}^{{}^{\prime }}$, and (f) $x_{\mathit{latent}}^{{}^{\prime }}$.

Figure 12

The variation of classification performance accuracy using GLCM feature descriptor for different values of d with variation in θ as 0°, 45°, 90°, and 135°, (a) Offset matrix, and (b) accuracy versus d plot.

Figure 13

Sensitivity analysis of HOG parameters for radiogenomic classification: (a) number of bins, (b) block size, and (c) cell size.

Figure 14

The sensitivity analysis of LBP parameters: (a) the number of neighbors, and (b) the radius.

Figure 15

Comparison of feature extraction techniques for their feature set size and the resulting classification accuracy with RA using a $1$-NN classifier.

Figure 16

Schematic diagram of multi-omics HFS formation for the proposed classification framework.

Figure 17

Effect of RA on time required for (a) deep learning feature extraction time (minutes), (b) ML training time and (c) ML testing time for radiogenomic classification using BraTS-2021 dataset and 1-NN classifier (DLFET stands for deep learning feature extraction time and MLTrgT stands for Machine Learning Training Time).

Figure 18

Comparison of (a) feature extraction times for potential feature extraction techniques with (b) individual feature set size used in the proposed framework, and (c) classification time without preprocessing and feature extraction with RA using BraTS-2021 dataset and 1-NN classifier.