No abstract

Glioblastoma (GBM) is a heterogeneous and lethal brain cancer. These tumors are followed using magnetic resonance imaging (MRI), which is unable to precisely identify tumor cell invasion, impairing effective surgery and radiation planning. We present a novel hybrid model, based on multiparametric intensities, which combines machine learning (ML) with a mechanistic model of tumor growth to provide spatially resolved tumor cell density predictions. The ML component is an imaging data-driven graph-based semi-supervised learning model and we use the Proliferation-Invasion (PI) mechanistic tumor growth model. We thus refer to the hybrid model as the ML-PI model. The hybrid model was trained using 82 image-localized biopsies from 18 primary GBM patients with pre-operative MRI using a leave-one-patient-out cross validation framework. A Relief algorithm was developed to quantify relative contributions from the data sources. The ML-PI model statistically significantly outperformed (p < 0.001) both individual models, ML and PI, achieving a mean absolute predicted error (MAPE) of 0.106 ± 0.125 versus 0.199 ± 0.186 (ML) and 0.227 ± 0.215 (PI), respectively. Associated Pearson correlation coefficients for ML-PI, ML, and PI were 0.838, 0.518, and 0.437, respectively. The Relief algorithm showed the PI model had the greatest contribution to the result, emphasizing the importance of the hybrid model in achieving the high accuracy.

BACKGROUND: MRI-based modeling of tumor cell density (TCD) can significantly improve targeted treatment of Glioblastoma (GBM). Unfortunately, interpatient variability limits the predictive ability of many modeling approaches. We present a Transfer Learning (TL) method that generates individualized patient models, grounded in the wealth of population data, while also detecting and adjusting for interpatient variabilities based on each patient’s own histologic data. METHODS: We recruited primary GBM patients undergoing image-guided biopsies and preoperative imaging including contrast-enhanced MRI (CE-MRI), Dynamic-Susceptibility-Contrast (DSC)-MRI, and Diffusion Tensor Imaging (DTI). We calculated relative cerebral blood volume (rCBV) from DSC-MRI and mean diffusivity (MD) and fractional anisotropy (FA) from DTI. Following image coregistration, we assessed TCD for each biopsy and identified corresponding localized MRI measurements. We then explored a range of univariate and multivariate predictive models of TCD based on MRI measurements in a generalized one-model-fits-all (OMFA) approach. We then implemented both univariate and multivariate individualized TL predictive models, which harness the available population level data but allow for individual variability in their predictions. Finally, we compared Pearson correlation coefficients and mean absolute error between the individualized TL versus generalized OMFA models. RESULTS: TCD significantly correlated with rCBV (r=0.33,p<0.0001) and T1+C (r=0.36,p<0.0001) on univariate analysis after correcting for multiple comparisons. With single variable modeling (using rCBV), TL increased predictive performance (r=0.53, MAE=15.19%) compared to OMFA (r=0.27, MAE=17.79%). With multivariate modeling, TL further improved performance (r=0.88, MAE 5.66%) compared to OMFA (r=0.39, MAE=16.55%). CONCLUSION: TL significantly improves predictive modeling performance for quantifying tumor cell density in GBM.

space as a transformation space to detect radially symmetric blobs. SIFT 18 , SURF 19 and BRISK 20 are region detectors. Each of the region detectors extracts scale invariant features to detect small objects but may suffer from poor performance in optical imaging 21. Recently, the Laplacian of Gaussian (LoG) detector 22,23 , from scale space theory, has attracted attention in blob detection 8,24. Similar to the radially symmetric detector, the LoG detector is unreliable in detecting rotationally asymmetric blobs. To solve this, LoG extensions have been proposed, including the Difference of Gaussian (DoG) 18,25-27 and the Generalized Laplacian of Gaussian (gLoG) 21. While each approach detects small blobs to some extent, non-blob objects are detected as false blob candidates resulting in over-detection. A post-pruning procedure can remove false blob candidates, but results have been inconsistent 28. Here we focus on detecting individual glomeruli in MR images of the kidney as a specific blob detection problem. To date, most biomarkers of kidney pathology have come from histology using destructive techniques that estimate glomerular number 29-31. A non-destructive imaging approach to measuring nephron endowment provides a new marker for renal health and susceptibility to kidney disease. Cationic ferritin enhanced MRI (CFE-MRI) enables the detection of glomeruli in animals 32,33 and in human kidneys 5,32-34. Because each glomerulus is associated with a nephron, CFE-MRI may provide an important imaging marker to detect changes in the number of nephrons and susceptibility to renal and cardiovascular disease 5. Glomerulus detection by CFE-MRI presents difficulties because glomeruli are small and have a spatial frequency similar to image noise. Zhang et al. developed the Hessian-based Laplacian of Gaussian (HLoG) detector 1 and the Hessian-based Difference of Gaussian (HDoG) detector 28 to automatically detect glomeruli in CFE-MR images. They employed the LoG or DoG to smooth the images, followed by Hessian analysis of each voxel for pre-segmentation. Since LoG and DoG suffer from over-detection, a Variational Bayesian Gaussian Mixture Model (VBGMM) was implemented as a final step. LoG and DoG were the first two detectors applied to MR images of the kidney to identify glomeruli. However, deriving Hessian-based features from each blob candidate is computationally expensive, limiting high-throughput studies. In addition, unsupervised learning using the VBGMM in the post-pruning procedure requires a number of carefully tuned parameters for optimal clustering. Here we propose a new approach, termed UH-DoG, which applies joint constraints from spatial probability maps derived from U-Net, a deep learning model, and Hessian convexity maps derived from Hessian analysis on the DoG detector. The theoretical foundation of Hessian analysis guarantees that pre-segmentation will recognize all true convex blobs and some non-blob convex objects, resulting in a blob superset. Joining probability maps allows us to distinguish true blobs ...