Voxel-Wise GBM Recurrence Prediction Based on Sparse Attention Multi-Modal MR Image Fusion Coupling with Stem Cell Niches Proximity Estimation

Abstract

<h3>Purpose/Objective(s)</h3> Nearly all glioblastoma multiforme (GBM) patients experience local recurrence despite aggressive therapies. The outcomes of patients with recurrence continue to be disappointing, emphasizing the importance of early and voxel-wise recurrence prediction for early treatment intervention. Emerging evidence has linked the tumorigenesis of recurrence to stem cell niches (SCN). For the first time, our group introduced a novel multidimensional support vector machine (SVM) method coupling with SCN proximity estimation to predict GBM recurrence on the voxel level. The current study aims to develop an L1-sparsity regularized attention fusion U-net (L1-AF U-net) to further distinguish the most salient and complementary features from multi-modal MR images, improving voxel-wise GBM recurrence prediction. <h3>Materials/Methods</h3> With IRB approval, fifty patients with pre-and post-surgery MR scans were retrospectively solicited, including T1, T2, T1_CE, ADC, and FLAIR. Post-surgery MR scans included two months before the clinical diagnosis of GBM recurrence and the day of the radiologically confirmed recurrence. The recurrences were manually annotated on the contrast-enhanced T1 MR, which were deformably registered to the 2-month pre-recurrence MR scans. The processing pipeline included a sparse attention learning module with L1 norm regularization for multi-modal feature fusion and a U-net for voxel-wise GBM recurrence prediction. The L1 sparsity regularized attention network embedded in the U-net architecture learns the coherent and mutual-complementary features across five MR modalities while eliminating the inherent redundancy and noisy artifacts. The high-risk recurrence (HRR) proximity map was determined by the weighted sum of inverse distances from two possible origins of recurrence: SCN and tumor cavity. The U-net was trained given the discriminative feature fusion in HRR for voxel-wise GBM recurrence prediction <h3>Results</h3> On the 2-month pre-recurrence testing MRs, the U-net segmentation with sparse multi-modal attention fusion achieved a recall of 0.83±0.15, a precision of 0.75±0.12, and an F1-score of 0.79±0.14. As a comparison, our previous GBM recurrence prediction using proximity estimation coupled SVM (SVM_PE) achieved a recall of 0.80±0.10, a precision of 0.69±0.14, and an F1-score of 0.74±0.10. A Student's t-test has also been performed to evaluate the significance of the proposed method over SVM_PE. The P-values are 0.0013 (for recall), 4.97 × 10⁻¹⁰ (for precision) and 7.94 × 10⁻⁸ (for F1-score), respectively. <h3>Conclusion</h3> We present an advanced U-net-based GBM recurrence prediction framework with sparse multi-modal MR fusion. This constrained attention fusion of the multi-modal MR images facilitated the U-net to learn more effective features from the multi-modal MR images and enhanced the prediction performance comparing to SVM method.