Wang, Lujia

Abstract Intra-tumor genetic heterogeneity is an important cause of treatment failure of GBM. Using MRI and image-localized biopsies, it is possible to train machine learning (ML) models to predict regional genetic status. However, biopsy samples are limited, making it difficult to train a robust ML model. We proposed a data-inclusive model called Weakly Supervised Ordinal Support Vector Machines (WSO-SVM) which leverages the vast amount of MRI data outside the sparsely sampled biopsy regions to augment the biopsy samples to improve ML accuracy. Our study included a unique dataset of 104 image-localized biopsies with spatially matched multiparametric MRI from 30 untreated Glioblastoma (GBM) patients. Each biopsy sample went through genetic sequencing analysis and our study focused on two GBM hallmark genes, EGFR and PDGFRA. For each gene, a biopsy sample was labeled as “altered” if the copy number of this gene was amplified or a mutation was found, and “non-altered” otherwise. From the localized region of six MRI contrasts from T1gd, T2w, diffusion and perfusion imaging, over 300 texture features were extracted. To account for biopsy sample location uncertainty, six neighboring regions of the biopsy sample including four neighbors two pixels away from the biopsy location and two neighbors on adjacent slices were also included in model training. WSO-SVM achieved 0.83 accuracy, 0.77 sensitivity, and 0.86 specificity for classifying EGFR; 0.77 accuracy, 0.74 sensitivity, and 0.79 specificity for classifying PDGFRA, based on 10-fold cross validation. Furthermore, using the trained models, we generated regional EGFR and PDGFRA alteration maps for each patient within the enhancing and non-enhancing tumoral areas. On average we found a greater proportion of enhancing tumor with co-alteration than non-enhancing tumor, while this trend was reversed when considering only one altered gene. The ratio of EGFR vs PDGFRA alterations was higher in non-enhancing tumor than enhancing tumor.

Abstract Recurrent glioblastoma is challenging to distinguish from so called “treatment effect” on routine clinical imaging. Further, within tumor heterogeneity reveals that some regions can be histologically dominated by tumor progression whilst others can be dominated by secondary effects of treatment response. Apparent tumor progression on MRI can be very difficult to manage clinically as it is unclear the degree to which the imaging changes are actually tumor progression vs response to treatment (including inflammatory response and necrosis). In this analysis, we study a unique cohort of patients for whom image localized-biopsies reveal heterogeneity in response vs progression. Our dataset included 70 biopsy samples from 32 patients with GBM each histolopatholgically characterized for tumor abundance vs immune infiltrate. Six multiparametric MRI contrasts were available, including T1, T1gd, T2, FLAIR, SWI, and ADC. Images were co-registered. Radiomic (statistical + texture) features were extracted from the region of six image contrasts locally matched with each biopsy sample. Machine learning models were built to predict each biomarker using radiomic features. Leave-one-out cross validation was used to evaluate the prediction accuracy. Radiomic features were found to be informative to the prediction of biomarkers. ANOVA tests show significant improvement of using radiomic features compared with the null model. The prediction accuracy was higher when considering the biomarkers on a binary scale using the median as the cutoff than on a numerical scale. Spatially-informed radiomics models for tumor progression vs treatment effect are possible and can play an instrumental role in navigating confounding imaging changed common during treatment progression.

Abstract BACKGROUND An important challenge in radiomics research is reproducibility. Images are collected on different image scanners and protocols, which introduces significant variability even for the same type of image across institutions. In the present proof-of-concept study, we address the reproducibility issue by using domain adaptation – an algorithm that transforms the radiomic features of each new patient to align with the distribution of features formed by the patient samples in a training set. METHOD Our dataset included 18 patients in training with a total of 82 biopsy sample. The pathological tumor cell density was available for each sample. Radiomic (statistical + texture) features were extracted from the region of six image contrasts locally matched with each biopsy sample. A Gaussian Process (GP) classifier was built to predict tumor cell density using radiomic features. Another 6 patients were used to test the training model. These patients had a total of 31 biopsy samples. The images of each test patient were purposely normalized using a different approach, i.e., using the CSF instead of the whole brain as the reference. This was to mimic the practical scenario of image source discrepancy between different patients. Domain adaptation was applied to each test patient. RESULTS Among the 18 training patients, the leave-one-patient-out cross validation accuracy is 0.81 AUC, 0.78 sensitivity, and 0.83 specificity. When the trained model was applied to the 6 test patients (purposely normalized using a different approach than that of the training data), the accuracy dramatically reduced to 0.39 AUC, 0.08 sensitivity, and 0.61 specificity. After using domain adaption, the accuracy improved to 0.68 AUC, 0.62 sensitivity, and 0.72 specificity. CONCLUSION We provide candidate enabling tools to address reproducibility in radiomics models by using domain adaption algorithms to account for discrepancy of the images between different patients.

Abstract INTRODUCTION The quantification of intratumoral heterogeneity – through radiomics-based approaches - can help resolve the regionally distinct genetic drug targets that may co-exist within a single Glioblastoma (GBM) tumor. While this offers potential diagnostic value under the paradigm of individualized oncology, clinical decision-making must also consider the degree of uncertainty associated with each model. In this study, we evaluate the performance of a novel machine-learning (ML) algorithm, called Gaussian Process (GP) modeling, that can quantify the impact of multiple sources of uncertainty in ML model development and prediction accuracy, including variabilities in the copy number measurement, radiomics features, training sample characteristics, and training sample size. METHOD We collected 95 image-localized biopsies from 25 primary GBM patients. We coregistered stereotactic locations with preoperative multi-parametric MRI features (conventional MRI, DSC perfusion, Diffusion Tensor Imaging) to generate spatially matched pairs of MRI and copy number variants (CNV) for for each biopsy. We developed a Gaussian Process (GP) model to predict CNV for Epidermal Growth Factor Receptor (EGFR) based on MRI radiomic features in each patient. We used leave-one-patient-out cross validation to quantify prediction accuracy and model uncertainty. Spatial prediction and uncertainty (p-value) maps were overlaid to help visualize regional genetic variation of EGFR and uncertainty of the radiomic predictions. RESULT: The initial GP radiomics model for EGFR amplification (CNV > 3.5) produced a sensitivity of 0.8 and specificity of 0.8. Samples/regions associated with high uncertainty (p-value >0.05) correlated with either 1) extrapolation of radiomic features from the training set-defined feature space or 2) insufficient training samples in the feature space. CONCLUSION We present a ML-based model that quantifies spatial genetic heterogeneity in GBM, while also estimating model uncertainties that result from multi-source data variabilities. This approach lays the groundwork for prospective clinical integration of modeling-based diagnostic approaches in the paradigm of individualized medicine.