Conte, Gian Marco

Abstract Clinical management of IDH wildtype glioblastoma (GBM) and tumefactive multiple sclerosis (tMS) is drastically different. GBM requires maximal safe resection followed by chemoradiation, while tMS outcome is worsened by surgery and radiotherapy. Misdiagnosis of brain lesions may expose patients to unnecessary anxiety, surgery, or radiotherapy. Noninvasive methods to help with accurate diagnosis of tumor and non-tumor etiologies are needed. tMS subjects were identified by the Mayo Clinic Center for Multiple Sclerosis and Autoimmune Neurology (CMSAN). GBM subjects were matched to tMS by age at diagnosis, index MRI date, and sex. Inclusion criteria included one cm minimal lesion size and pre-operative post-contrast T1 and T2 images available for analysis. To identify potential experimental design bias, MRI technology and clinical characteristics were compared across groups. A 3D-DenseNet121 was used to develop a classification model using prespecified parameters: 600 epochs, batch size 16, learning rate 10-3, cross entropy loss, and AdamW optimizer. The stopping rule was defined as three sequential epoch cross entropy loss < 0.02. Five-fold cross validation (CV) was utilized to obtain an unbiased estimate of predictive performance (overfitting) and sensitivity of the 3D-DenseNet121 algorithm was evaluated. 224 subjects (114 GBM, 110 tMS) were analyzed: 40% male, median age 46. After one round of CV, the average unbiased estimate of sensitivity and specificity for predicting GBM from the final classification model was 0.86 and 0.73, respectively. To assess the reproducibility of the deep learning algorithm (3D-DenseNet121), the exact same CV folds were rerun resulting in average sensitivity and specificity of 0.82 and 0.75, respectively. Obtaining unbiased estimates of model performance using internal validation is critical before evaluating a final model on an external dataset. Understanding sensitivity of a deep learning algorithm is important for model validity. Differentiation of GBM and tMS appears reasonable with MRI; external validation is in progress.

Abstract INTRODUCTION Prior to 2021, (classic) glioblastoma (GBM) required absence of IDH mutation and histologic features of hypercellularity, nuclear atypia, microvascular proliferation (MVP), and necrosis. Absence of histologic features can be a result of undersampling, with molecular, radiologic, and clinical features consistent with GBM. Accordingly, WHO 2021 defined molecular glioblastoma as diffuse glioma IDHwt, H3wt, with one or more of MVP, necrosis, TERT promoter mutation, EGFR gene amplification (EGFRamp), and/or +7/−10. We hypothesized that there is clinical and radiographic heterogeneity among tumors defined as molecular GBM. METHODS We identified 121 patients that met the criteria for molecular GBM. 121 Classic GBM, defined as IDHwt with hypercellularity, MVP or necrosis, were matched to molecular GBM cases by age, sex, and diagnosis date. All cases were reviewed for relevant histologic features and molecular features. Examined clinical variables included: first symptom, time from first symptom to date of pathological diagnosis, date of first MRI, tumor location, radiographic features of enhancement, diffusion or mass effect, time to progression and overall survival. Clinical, radiologic, and molecular features were compared between classic and molecular GBM. RESULTS Currently,160 cases have been reviewed. Preliminary results show that there is a longer time from first symptom to diagnosis for molecular GBM. Molecular GBM is significantly more likely to be discovered coincidentally, and without enhancement or radiographic necrosis. Classic GBM is significantly more likely to present with focal neurologic deficit as first symptom, regional mass effect on index imaging and have monosomy 10. CONCLUSION While pathologic undersampling accounts for a significant proportion of molecular GBM, our data suggest that there are significant clinical, radiographic, and outcome differences between classical and molecular GBM. Final analysis of the entire cohort will help better define this clinical spectrum.

Abstract There have been several attempts to predict the MGMT promoter methylation status of gliomas using deep learning (DL), with mixed results. These studies did not account for IDH mutation status and did not utilize a uniform methylation assay, introducing potential confounders in the experimental design. We assessed the performance of DL applied to brain MRI to predict the MGMT promoter methylation status of IDH-wildtype gliomas tested with a standardized assay. We included 539 IDH-wildtype gliomas tested with real-time methylation-specific PCR from 2017-2022; 205 (38%) were MGMT promoter methylated. The assay targets 8 downstream MGMT CpG sites using ACTB as control; methylation status is determined based on the difference of PCR crossing-point values: methylated if difference < 5, unmethylated if difference > 7, indeterminate otherwise. We used post-contrast T1 and T2 MRIs as inputs for our models trained using stratified 5-fold cross-validation, 400 epochs, batch size of 16, a learning rate scheduler starting at 10-3, weighted binary cross-entropy loss, and AdamW optimizer. We repeated the experiments using two 3D DL architectures (DenseNet121 and EfficientNetB0) and two image preprocessing pipelines: the BraTS protocol and one that uses padding instead of registration to an MRI atlas, keeping constant other experimental variables. For Densenet121, the best cross-validation AUROCC and F1-score (median, median absolute deviation) were 0.67 (0.06), 0.59 (0.02) using the BraTS protocol, and 0.67 (0.06), 0.58 (0.01) using padding. For EfficientNetB0, were 0.67 (0.02), 0.66 (0.01) using the BraTS protocol and 0.67 (0.01), 0.68 (0.01) using padding. Our findings suggest that it is unlikely that DL models trained on brain MRI can predict the MGMT promoter methylation of IDH-wildtype gliomas. Importantly, our study is the first to focus on an IDH-wildtype adult diffuse glioma cohort tested with a single, standardized MGMT methylation assay, eliminating the variability introduced by different assays and IDH status.

Abstract Incomplete MRI exams reduce the possibility of applying AI models. We previously trained two generative adversarial networks (GANs) on 135 subjects from the BraTS-2017 dataset to generate T1 from post-contrast T1 and FLAIR from T2 sequences (baseline models). To test the impact of the training set size, here, we trained the same models using 1251 subjects from the BraTS-2021 dataset. We used the same architecture and hyperparameters from the baseline models: batch size of 1, learning rate 2x10-4, Adam optimizer, cross-entropy and L1-loss. We trained two versions of the updated models: one stopped at an early checkpoint (early models) and one after 50 epochs (late models). We tested all models on institutional MRIs from 487 IDH-wt gliomas. Synthetic MRIs were compared to the originals using the structural-similarity-index (SSI) and MSE. We simulated scenarios where either the T1, FLAIR, or both were missing and used their synthetic version as inputs of a segmentation model. We compared the segmentations using the dice score (DSC) and used Friedman and Dunn’s test to compare the scores and correct for multiple comparisons. Median SSI for the T1 were 0.957, 0.947, 0.947, and median MSE were 0.006, 0.014, 0.008 for the baseline, early, and late models. For the FLAIR, median SSI were 0.924, 0.908, 0.915, and median MSE were 0.016, 0.016, 0.019 for the baseline, early, and late models. The DSC ranges for the baseline and updated models were (0.655 - 0.953), (0.420 - 0.952), and (0.610 - 0.952). The baseline and late models showed no statistically significant differences in DSC, but both performed significantly better than the early models. We show that GANs trained to synthesize brain MRIs on a small cohort perform similarly to those trained on an x10-larger cohort, making them a viable option for rare diseases or institutions with limited resources.

Abstract Missing MRI sequences are common in radiology due to differences in protocols, time constraints, and image artifacts. Missing sequences are a nuisance for subsequent algorithms, such as automated segmentation algorithms requiring multiple MRI sequences (e.g., HD-GLIO, which requires T2, T1, post-contrast T1 (T1c), and FLAIR). Accurate tumor segmentation is a fundamental step in extracting radiomics analysis to use, for example, to predict the IDH status of glioma or their overall survival. We validated a conditional (pix2pix) and super-resolution GAN (SOUP-GAN) for imputing MRI sequences on 285 adult glioma with complete MRI sequences available. We simulated missing T1 sequences and subsequently imputed them from T1c using pix2pix and from sagittal T1 using SOUP-GAN. Using HD-GLIO, we obtained the true segmentations from the complete MRI sequences. HD-GLIO was also used after replacing the original T1 sequence with the one imputed with pix2pix and SOUP-GAN. Thus, there were three sets of segmentations for each subject. We compared T2 and T1c radiomics features extracted with Pyradiomics using segmentations obtained with complete data with features extracted from segmentation obtained with the imputed T1 sequences using intraclass correlation coefficient (ICC) and Bland-Altman (BA) analysis. Median (interquartile range) ICC for shape features extracted from T2 and T1c were 0.82(0.76-0.92) using pix2pix and 0.89(0.86-0.94) using SOUP-GAN. Median ICC for first-order-statistics features were 0.90(0.83-0.93) for T2 and 0.90(0.88-0.92) for T1c using pix2pix, and 0.61(0.22-0.82) for T2 and 0.42(0.34-0.57) for T1c using SOUP-GAN. Median ICC for texture features were 0.86(0.82-0.87) for T2 and 0.87(0.79-0.90) for T1c using pix2pix, and 0.69(0.47-0.81) for T2 and 0.76(0.65-0.83) for T1c using SOUP-GAN. We show that T2 and T1c shape features (e.g., volumetric analysis) can be adequately extracted from segmentations obtained with an algorithm requiring multiple MRI sequences using T1 sequences imputed with GANs; however, the effect of imputation differed for first-order-statistics and texture features.

Abstract GBM is the most common primary malignant brain tumor in adults. O6-methylguanine DNA methyltransferase (MGMT) promoter methylation status is an important prognostic biomarker that predicts the response to temozolomide and guides the treatment decisions. At present, the only reliable way to determine MGMT promoter methylation status is through the analysis of tumor tissues. Given the limitations and complications of histology-based methods, an imaging-based approach for non-invasively MGMT promoter methylation status prediction is beneficial. This study aimed to develop and compare three different deep-learning-based approaches for predicting MGMT promoter methylation status non-invasively. We obtained 576 T2 weighted images (T2WIs) with their corresponding tumor masks and MGMT promoter methylation status from, The Brain Tumor Segmentation (BraTS) 2021 datasets. Dataset was split into five folds at the patient's level stratified by MGMT promoter methylation status to perform a 5-fold cross-validation. We developed three different models: voxel-wise, slice-wise, and whole-brain. For voxel-wise classification, methylated and unmethylated MGMT tumor masks were made into 1 and 2 with 0 background, respectively. We converted each T2WI into 32x32x32 patches. We trained a 3D-Vnet model for tumor segmentation. After inference, we constructed the whole brain volume based on the patch's coordinates. The final prediction of MGMT methylation status was made by majority voting between the predicted voxel values of the biggest connected component. For slice-wise classification, we trained an object detection model for tumor detection and MGMT methylation status prediction; then, we used majority voting for the final prediction. For the whole-brain approach, we trained a 3D Densenet121 for prediction. Whole-brain, slice-wise, voxel-wise, accuracy was 65.42%(SD 3.97%), 61.37%(SD 1.48%), and 56.84%(SD 4.38%) respectively. We found that across the whole-brain, slice-wise, and voxel-wise deep learning approaches, the whole-brain approach is the most effective approach for MGMT methylation status prediction on the BraTS 2021 dataset.

Abstract GBMs are highly aggressive tumors, evidenced by their rapid growth. Despite maximal treatment, median overall survival ranges between 16 and 20 months. Several studies use patients' diagnosis time(pre-surgical) imaging to predict their survival, but having a model that can predict the days till death for the patients who received the first standard step treatments (Stupp protocol) and now showing secondary progression signs is highly needed. In this study, we aimed to predict the overall survival of Glioblastoma(GBM) patients by developing a multimodal deep learning model using post-operative magnetic resonance imaging(MRI) and patients' age. We included GBM patients who met the following inclusion criteria: having maximum surgical resection followed by concurrent chemoradiation and showing secondary progression signs in follow-up imaging while being on adjuvant temozolomide treatment. We obtained contrast-enhanced T1(CET1) and T2-weighted(T2W) MRIs at progression time and patients' age and death dates. We calculated the days of survival from the day of the MRI. Dataset was split into train, validation, and test set at the patient level. CET1 and T2W MRIs were skull-stripped and coregistered for each patient. To mitigate data imbalance, we used the deep imbalance regression method for training a 3-D Densenet121 model with the concatenated CET1 and T2W as inputs. One hundred features were extracted from the optimized model and concatenated with the normalized age of each patient; the resulting output vectors were passed through the second dense neural network for final prediction. We included 110 patients who met the inclusion criteria. With the overall survival ranging from 16 to 4500 days, a mean of 582.4 and a standard deviation of 694.3, we achieved the test set's root mean square error of 261 days. Further refinement with more data, including clinical data such as comorbidities and external validation, is required prior to adoption in clinical practice.

Abstract INTRODUCTION Glioblastomas (GBMs) are highly aggressive tumors. A common clinical challenge after standard of care treatment is differentiating tumor progression from treatment-related changes, also known as pseudoprogression (PsP). Usually, PsP resolves or stabilizes without further treatment or a course of steroids, whereas true progression (TP) requires more aggressive management. Differentiating PsP from TP will affect the patient's outcome. This study investigated using deep learning to distinguish PsP MRI features from progressive disease. METHOD We included newly diagnosed GBM patients who met the inclusion criteria, including a new or increasing enhancing lesion size within the original radiation field and who had clinical, medication, and any histopathology available. The interpretation of the MR images at the image inclusion time point had to be indeterminate. We labeled those who subsequently were stable or improved on imaging and clinically as PSP and those with clinical and imaging deterioration as TP. A subset of subjects underwent a second resection. We labeled these subjects as PSP or TP based on the histological diagnosis. We performed skull stripping, coregistered contrast-enhanced T1 MRIs with T2-weighted images for each patient and used them as input to a 3-D Densenet121 model. We used several augmentation techniques and five-fold cross-validation to achieve more robust predictions. RESULT We included 124 patients who met the criteria, and of those, 63 were PsP and 61 were TP. We trained a deep learning model that achieved 76.4% (ranged 70%-84%, SD 5.122) mean accuracy over the 5 folds, 0.7560(ranged 0.6553-0.8535, SD 0.069) mean AUROCC, 88.72% (SD 6.86) mean sensitivity, and 62.05%(SD 9.11) mean specificity. CONCLUSION We report the development of a deep learning model that distinguishes PsP from TP in GBM patients treated per the Stupp protocol. Further refinement and external validation are required prior to widespread adoption in clinical practice.

Abstract PURPOSE Though patients with glioblastoma generally have poor overall survival (OS), OS estimates range significantly. We sought to develop a radiomics-based model predictive of OS in glioblastoma using preoperative MRI scans to identify these differences and guide clinical care. METHODS The primary endpoint was OS, determined from the date of surgery to death or last follow-up. T1, FLAIR, T2, and post-contrast T1 (T1c) MRI sequences were obtained for all patients. Using HD-GLIO-AUTO, lesion segmentation yielded three masks corresponding to the regions of contrast enhancement, non-enhancement, and total volume. Radiomics features were extracted from each region from the T2 and T1c sequences independently (i.e. either from T2 or T1c) and in combination. Age at surgery and gender were also considered. The data was split into training (75%) and test (25%) sets, stratifying for age at surgery. Kaplan-Meier analysis and the Coxnet model were used. For Coxnet, Elastic Net was used as the penalty, with optimization of the L1 coefficient using 5-fold cross-validation and evaluation via the concordance index (C-index). RESULTS From 3 institutions, 383 patients were included. Median OS was 395 days: 346 patients died, and 37 were right-censored. Age > 65 years was predictive of decreased OS (p < 0.005), but gender was not (p=0.16). For the Coxnet model, the C-index was 0.67 (standard deviation: 0.01) considering T2 and T1c sequences for the total lesion volume, with a C-index of 0.65 for the test set. For the enhancing volume, the model had a C-index of 0.71 for the test set using T1c sequences. Overall, C-indices ranged from 0.62-0.71. CONCLUSIONS We present a novel radiomics-based model predictive of OS in glioblastoma across a multi-institutional dataset. Ongoing efforts will incorporate immediate postoperative imaging and IDH and MGMT status into the proposed model to further improve prognostic ability and help guide clinical decision-making.