Conte, Gian Marco

Abstract INTRODUCTION Glioblastomas (GBMs) are highly aggressive tumors. Despite multimodal treatment, its median overall survival ranges between 16 and 20 months. The standard treatment regimen consists of surgical resection followed by concurrent chemoradiotherapy and adjuvant temozolomide. Despite temozolomide’s effectiveness, it may cause the clinical challenge of treatment-related progression also known as pseudoprogression(PsP). Usually, PSP resolves or stabilizes without further treatment, whereas a true progression (TP) requires more aggressive management. Identifying PSP from TP will affect the patient’s treatment plan. Conventional magnetic resonance imaging (MRI) reading techniques cannot distinguish these entities. This study investigated the feasibility of using deep learning to distinguish PsP from TP. METHOD We included GBM patients who met our inclusion criteria. We evaluated all cases to see if they had a new enhancing lesion within the original radiation field or an increase in the size of an existing lesion. The challenging MRIs were collected. Clinical notes regarding tumor and recurrence location, clinical history, and medication were collected. We labeled the ones who stayed stable or improved in the imaging and clinical situation as PSP and those with further imaging and clinical deterioration as TP. We coregistered Contrast-enhanced-T1 MRIs with T2-weighted images for each patient. We performed five-fold cross-validation to generalize the performance. We trained A 3-D Densenet121 model to establish the prediction. We selected the best models with the highest accuracy. RESULT After reviewing 1000 patients, we included 124 patients whose imaging showed suspicious progression and their medicational histories were completely retrievable; 63 PsP, and 61 TP. We developed a deep learning model based on the whole dataset. The 5-fold cross-validation revealed that the mean area under the curve (AUC) was 0.81. CONCLUSION We report the development of a deep learning model that diagnoses PsP from TP in patients who received temozolomide. Further refinement and external validation are required prior to widespread adoption in clinical practice.

There is a growing demand for high-resolution (HR) medical images for both clinical and research applications. Image quality is inevitably traded off with acquisition time, which in turn impacts patient comfort, examination costs, dose, and motion-induced artifacts. For many image-based tasks, increasing the apparent spatial resolution in the perpendicular plane to produce multi-planar reformats or 3D images is commonly used. Single-image super-resolution (SR) is a promising technique to provide HR images based on deep learning to increase the resolution of a 2D image, but there are few reports on 3D SR. Further, perceptual loss is proposed in the literature to better capture the textural details and edges versus pixel-wise loss functions, by comparing the semantic distances in the high-dimensional feature space of a pre-trained 2D network (e.g., VGG). However, it is not clear how one should generalize it to 3D medical images, and the attendant implications are unclear. In this paper, we propose a framework called SOUP-GAN: Super-resolution Optimized Using Perceptual-tuned Generative Adversarial Network (GAN), in order to produce thinner slices (e.g., higher resolution in the ‘Z’ plane) with anti-aliasing and deblurring. The proposed method outperforms other conventional resolution-enhancement methods and previous SR work on medical images based on both qualitative and quantitative comparisons. Moreover, we examine the model in terms of its generalization for arbitrarily user-selected SR ratios and imaging modalities. Our model shows promise as a novel 3D SR interpolation technique, providing potential applications for both clinical and research applications.

Abstract Primary sclerosis cholangitis (PSC) predisposes individuals to liver failure, but it is challenging for radiologists examining radiologic images to predict which patients with PSC will ultimately develop liver failure. Motivated by algebraic topology, a topological data analysis - inspired framework was adopted in the study of the imaging pattern between the “Early Decompensation” and “Not Early” groups. The results demonstrate that the proposed methodology discriminates “Early Decompensation” and “Not Early” groups. Our study is the first attempt to provide a topological representation-based method into early hepatic decompensation and not early groups.

Abstract Abnormalities on brain MRI may be identified in ~15% of individuals aged 50-66. It can be difficult to discriminate glioma, CNS lymphoma, Inflammatory Demyelinating Disease (CNSIDD) and solitary metastasis, and misdiagnosis may expose patients to unnecessary anxiety, surgery, or radiotherapy. CNS lymphoma requires only biopsy, solitary metastasis may be resected and radiated or radiated empirically, and high-grade glioma requires maximal safe resection followed by chemoradiation. CNSIDD should only be biopsied when diagnostic uncertainty requires it, and resection and radiotherapy are unnecessary, introducing unwarranted morbidity. Polygenic risk models can identify patients at the highest risk of developing glioma; we hypothesized that these polygenic models could help with differential diagnosis of indeterminate brain lesions. We also hypothesized that race would be an important contributing factor in the models. In the initial discovery and validation European (EUR) cohorts the mean probability of glioma for IDHmut non-codeleted glioma was 0.55 and 0.52, respectively, and in healthy controls was 0.19 and 0.21, respectively. To further evaluate sensitivity, we analyzed additional genotype data from 867 gliomas (764 EUR, 54 AFR, 24 AMR, 15 EAS, 10 SAS) from The Cancer Genome Atlas (TCGA). Across 764 EUR IDHmut non-codeleted glioma, the mean probability 0.53, whereas across 54 AFR and 24 AMR the mean was 0.22 and 0.32, respectively. To evaluate specificity, we analyzed 3200 TCGA patients with primary tumor types that commonly metastasize to the brain (2676 EUR, 365 AFR, 46 AMR, 95 EAS, 18 SAS), and 236 AFR healthy controls. For patients with non-CNS primary tumors, the mean probability ranged from 0.09-0.18 for EUR and 0.04-0.07 for AFR. For AFR healthy controls, the mean was 0.05. Overall, race is a significant factor for polygenic risk models. Further work entails evaluating polygenic risk models in IDHwt glioma, CNS lymphoma and CNSIDD cohorts as well as in additional races.

Abstract BACKGROUND Appropriate patient stratification is of paramount importance in patients with primary central nervous system lymphomas (PCNSL) in order to improve therapeutic choices and to reduce treatment-related neurotoxicity. Age and performance status, two widely used prognostic indicators, may not be the most appropriate parameters for an optimal patient stratification, as age seldom reveals the true biological frailty of a patient and performance status scores are subject to inter-rater variability. Quantitative muscle biomarkers such as the skeletal-muscle-index at the third lumbar vertebra (L3-SMI) and temporal muscle thickness (TMT) are associated with worse prognosis in several oncological diseases. We aim to evaluate the role of these biomarkers in predicting survival in patients with PCNSL undergoing high-dose methotrexate-based chemotherapy. MATERIAL AND METHODS L3-SMI and TMT were calculated on abdominal CT and brain high-resolution 3D-T1-weighted-MR images, respectively, using predefined validated methods. Standardized sex-specific cut-offs were used to divide patients in different risk categories. Kaplan-Meier plots were calculated, and survival analysis was performed using log-rank tests, univariate, and multivariable Cox-regression models, calculating hazard ratios (HR) and 95% confidence intervals (CI), also adjusting for potential confounders (age, sex, and performance status). RESULTS Forty-three patients were included in this study. Median follow-up was 23 months (interquartile range 12–40); at median follow-up, rates of progression-free and overall survival for the cohort were 46% and 57%, respectively. Thirteen (30%) and 11 (26%) patients showed L3-SMI or TMT values below the predefined cut-offs. Subgroup analyses showed a significant association between quantitative muscle mass biomarkers and progression-free and overall survival. One-year progression free and overall survival rates were 8% and 21% for the 13 patients with L3-SMI below the standard cut-off value, respectively, compared to 66% and 68% for the 30 patients with L3-SMI above the cut-off values. Likewise, one-year progression free and overall survival rates were 10% and 15% for the 11 patients with low TMT, respectively, compared to 61% and 70% for the 32 patients with high TMT. In Cox-regression multivariable analysis patients with low L3-SMI or TMT showed significantly worse progression-free (HR 4.40, 95%CI 1.66–11.61, p = 0.003; HR 4.40, 95%CI 1.68–11.49, p=0.003, respectively) and overall survival (HR 3.16, 95%CI 1.09–9.11, p = 0.034; HR 4.93, 95%CI 1.78–13.65, p=0.002, respectively) compared to patients with high L3-SMI or TMT. CONCLUSION Quantitative muscle mass evaluation assessed by both L3-SMI and TMT is a promising tool to identify PCNSL patients at high risk of negative outcome. Confirmatory studies on larger independent series are warranted.

Abstract Introduction Tumefactive multiple sclerosis (MS) and glioblastoma (GBM) can be difficult to differentiate. Clinical presentations are nonspecific and may occur in patients with both diseases. Magnetic resonance imaging (MRI) scans are routinely used to generate an initial differential diagnosis. Although imaging features can help distinguish these entities, there are no pathognomonic signs, neither on traditional nor advanced MRI. Moreover, the ability to differentiate these entities depends on the radiologists’ experience. Ensuring the correct diagnosis has important therapeutic and prognostic implications, as the management of these diseases is very different. To date, a brain biopsy often represents the only option to obtain a diagnosis, but it may be inconclusive and could add unnecessary morbidity. We hypothesize that a machine-learning tool based on routinely performed MRI scans could guarantee a more reliable, fast, and reproducible diagnosis. Methods Only patients with a histological diagnosis were included in the study. Patients that underwent any form of radiation therapy before the surgery were excluded to avoid confounding factors due to treatment-related effects. A total of 99 MS and 257 GBMs were included, and preoperative T2-weighted images were collected. For each patient, the lesions' segmentation was performed, including both the contrast-enhancing and non-enhancing areas of the lesions. The segmentation masks were used to extract radiomic features from the T2-weighted sequences. All MRI images were standardized to the same mean and standard deviation before feature extraction and resampled to a voxel size of 1.0 × 1.0 × 1.0 mm. Features extraction was performed using Pyradiomics (v. 3.0.1) and included first order, shape, and texture features, totaling 100 features. The classifier consisted of a support vector machine (SVM) model. Patients were randomly assigned to the training and test sets with an 80-20% proportion for each class. To select the best hyperparameters for the SVM classifier, a grid-search analysis was performed on the training set using a nested cross-validation scheme (k=5). All experiments were performed using Python (v. 3.5) and the Scikit-learn library (v. 0.23.1). Results The mean accuracy and area under the ROC curve for the best set of hyperparameters on the validation set were 0.83 and 0.86, respectively. The weighted-average precision, recall, and F1 scores on the test set were 0.78, 0.76, and 0.77, respectively. Conclusions Radiomic features extracted from a routinely acquired MRI sequence showed promising results in differentiating tumefactive multiple sclerosis from glioblastoma, opening the possibility to obtain a reliable diagnosis without the need for a brain biopsy. Future experiments Future experiments will focus on using post-contrast T1-weighted images for the radiomic features extractions, testing multiple machine learning models and features selection algorithms, and including more diagnostic classes (i.e., metastasis, lymphomas). Citation Format: Gian Marco Conte, Jeanette E. Eckel-Passow, W. Oliver Tobin, Paul Decker, Daniel H. Lachance, Robert B. Jenkins, Bradley J. Erickson. Differentiation of tumefactive multiple sclerosis and glioblastoma using radiomics features extracted from magnetic resonance imaging and machine learning [abstract]. In: Proceedings of the AACR Virtual Special Conference on Artificial Intelligence, Diagnosis, and Imaging; 2021 Jan 13-14. Philadelphia (PA): AACR; Clin Cancer Res 2021;27(5_Suppl):Abstract nr PO-023.

PurposeOrgan segmentation of computed tomography (CT) imaging is essential for radiotherapy treatment planning. Treatment planning requires segmentation not only of the affected tissue, but nearby healthy organs‐at‐risk, which is laborious and time‐consuming. We present a fully automated segmentation method based on the three‐dimensional (3D) U‐Net convolutional neural network (CNN) capable of whole abdomen and pelvis segmentation into 33 unique organ and tissue structures, including tissues that may be overlooked by other automated segmentation approaches such as adipose tissue, skeletal muscle, and connective tissue and vessels. Whole abdomen segmentation is capable of quantifying exposure beyond a handful of organs‐at‐risk to all tissues within the abdomen.MethodsSixty‐six (66) CT examinations of 64 individuals were included in the training and validation sets and 18 CT examinations from 16 individuals were included in the test set. All pixels in each examination were segmented by image analysts (with physician correction) and assigned one of 33 labels. Segmentation was performed with a 3D U‐Net variant architecture which included residual blocks, and model performance was quantified on 18 test cases. Human interobserver variability (using semiautomated segmentation) was also reported on two scans, and manual interobserver variability of three individuals was reported on one scan. Model performance was also compared to several of the best models reported in the literature for multiple organ segmentation.ResultsThe accuracy of the 3D U‐Net model ranges from a Dice coefficient of 0.95 in the liver, 0.93 in the kidneys, 0.79 in the pancreas, 0.69 in the adrenals, and 0.51 in the renal arteries. Model accuracy is within 5% of human segmentation in eight of 19 organs and within 10% accuracy in 13 of 19 organs.ConclusionsThe CNN approaches the accuracy of human tracers and on certain complex organs displays more consistent prediction than human tracers. Fully automated deep learning‐based segmentation of CT abdomen has the potential to improve both the speed and accuracy of radiotherapy dose prediction for organs‐at‐risk.