Inpainting arbitrary missing regions is challenging because learning valid features for various masked regions is nontrivial. Though U-shaped encoder-decoder frameworks have been witnessed to be successful, most of them share a common drawback of mask unawareness in feature extraction because all convolution windows (or regions), including those with various shapes of missing pixels, are treated equally and filtered with fixed learned kernels. To this end, we propose our novel mask-aware inpainting solution. Firstly, a Mask-Aware Dynamic Filtering (MADF) module is designed to effectively learn multi-scale features for missing regions in the encoding phase. Specifically, filters for each convolution window are generated from features of the corresponding region of the mask. The second fold of mask awareness is achieved by adopting Point-wise Normalization (PN) in our decoding phase, considering that statistical natures of features at masked points differentiate from those of unmasked points. The proposed PN can tackle this issue by dynamically assigning point-wise scaling factor and bias. Lastly, our model is designed to be an end-to-end cascaded refinement one. Supervision information such as reconstruction loss, perceptual loss and total variation loss is incrementally leveraged to boost the inpainting results from coarse to fine. Effectiveness of the proposed framework is validated both quantitatively and qualitatively via extensive experiments on three public datasets including Places2, CelebA and Paris StreetView.

Glioblastoma is characterized by diffuse infiltration into the surrounding tissue along white matter tracts. Identifying the invisible tumour invasion beyond focal lesion promises more effective treatment, which remains a significant challenge. It is increasingly accepted that glioblastoma could widely affect brain structure and function, and further lead to reorganization of neural connectivity. Quantifying neural connectivity in glioblastoma may provide a valuable tool for identifying tumour invasion. Here we propose an approach to systematically identify tumour invasion by quantifying the structural connectome in glioblastoma patients. We first recruit two independent prospective glioblastoma cohorts: the discovery cohort with 117 patients and validation cohort with 42 patients. Next, we use diffusion MRI of healthy subjects to construct tractography templates indicating white matter connection pathways between brain regions. Next, we construct fractional anisotropy skeletons from diffusion MRI using an improved voxel projection approach based on the tract-based spatial statistics, where the strengths of white matter connection and brain regions are estimated. To quantify the disrupted connectome, we calculate the deviation of the connectome strengths of patients from that of the age-matched healthy controls. We then categorize the disruption into regional disruptions on the basis of the relative location of connectome to focal lesions. We also characterize the topological properties of the patient connectome based on the graph theory. Finally, we investigate the clinical, cognitive and prognostic significance of connectome metrics using Pearson correlation test, mediation test and survival models. Our results show that the connectome disruptions in glioblastoma patients are widespread in the normal-appearing brain beyond focal lesions, associated with lower preoperative performance (P < 0.001), impaired cognitive function (P < 0.001) and worse survival (overall survival: hazard ratio = 1.46, P = 0.049; progression-free survival: hazard ratio = 1.49, P = 0.019). Additionally, these distant disruptions mediate the effect on topological alterations of the connectome (mediation effect: clustering coefficient −0.017, P < 0.001, characteristic path length 0.17, P = 0.008). Further, the preserved connectome in the normal-appearing brain demonstrates evidence of connectivity reorganization, where the increased neural connectivity is associated with better overall survival (log-rank P = 0.005). In conclusion, our connectome approach could reveal and quantify the glioblastoma invasion distant from the focal lesion and invisible on the conventional MRI. The structural disruptions in the normal-appearing brain were associated with the topological alteration of the brain and could indicate treatment target. Our approach promises to aid more accurate patient stratification and more precise treatment planning.

OBJECTIVE: The objective of this study was to characterize the abnormalities revealed by diffusion tensor imaging (DTI) using MR spectroscopy (MRS) and perfusion imaging, and to evaluate the prognostic value of a proposed quantitative measure of tumor invasiveness by combining contrast-enhancing (CE) and DTI abnormalities in patients with glioblastoma.METHODS: Eighty-four patients with glioblastoma were recruited preoperatively. DTI was decomposed into isotropic (p) and anisotropic (q) components. The relative cerebral blood volume (rCBV) was calculated from the dynamic susceptibility contrast imaging. Values of N-acetylaspartate, myoinositol, choline (Cho), lactate (Lac), and glutamate + glutamine (Glx) were measured from multivoxel MRS and normalized as ratios to creatine (Cr). Tumor regions of interest (ROIs) were manually segmented from the CE T1-weighted (CE-ROI) and DTI-q (q-ROI) maps. Perfusion and metabolic characteristics of these ROIs were measured and compared. The relative invasiveness coefficient (RIC) was calculated as a ratio of the characteristic radii of CE-ROI and q-ROI. The prognostic significance of RIC was tested using Kaplan-Meier and multivariate Cox regression analyses.RESULTS: The Cho/Cr, Lac/Cr, and Glx/Cr in q-ROI were significantly higher than CE-ROI (p = 0.004, p = 0.005, and p = 0.007, respectively). CE-ROI had significantly higher rCBV values than q-ROI (p < 0.001). A higher RIC was associated with worse survival in a multivariate overall survival (OS) model (hazard ratio [HR] 1.40, 95% confidence interval [CI] 1.06–1.85, p = 0.016) and progression-free survival (PFS) model (HR 1.55, 95% CI 1.16–2.07, p = 0.003). An RIC cutoff value of 0.89 significantly predicted shorter OS (median 384 vs 605 days, p = 0.002) and PFS (median 244 vs 406 days, p = 0.001).CONCLUSIONS: DTI-q abnormalities displayed higher tumor load and hypoxic signatures compared with CE abnormalities, whereas CE regions potentially represented the tumor proliferation edge. Integrating the extents of invasion visualized by DTI-q and CE images into clinical practice may lead to improved treatment efficacy.