HeeJoo Lim^1,2, Eunji Ha³, Suji Lee³, Sujung Yoon^3,4, In Kyoon Lyoo^3,4,5, and Taehoon Shin^1,2
1Division of Mechanical and Biomedical Engineering, Ewha Womans University, Seoul, Korea, Republic of, ²Graduate Program in Smart Factory, Ewha Womans University, Seoul, Korea, Republic of, ³Ewha Brain Institute, Ewha Womans University, Seoul, Korea, Republic of, ⁴Department of Brain and Cognitive Sciences, Ewha Womans University, Seoul, Korea, Republic of, ⁵Graduate School of Pharmaceutical Sciences, Ewha Womans University, Seoul, Korea, Republic of

Prediction of biological brain age is important as its deviation from chronological age can serve as a biomarker for degenerative neurological disorders. In this study, we suggest novel fusion deep learning algorithms which combine pre-engineered features and convolutional neural net (CNN) extracted features of T1-weighted MR images. Over all backbone CNN architectures, fusion models improved prediction accuracy (mean absolute error (MAE) = 3.40–3.52) compared with feature-engineered regression (MAE = 4.58–5.15) and image-based CNN (MAE = 3.60–3.95) alone. These results indicate that using both features derived from convolution and pre-engineering can complement each other in predicting brain age.

Grant Nikseresht¹, Ashish A. Tamhane², Carles Javierre-Petit³, Arnold M. Evia², David A. Bennett², Julie A. Schneider², Gady Agam¹, and Konstantinos Arfanakis^2,3
1Department of Computer Science, Illinois Institute of Technology, Chicago, IL, United States, ²Rush Alzheimer's Disease Center, Rush University Medical Center, Chicago, IL, United States, ³Department of Biomedical Engineering, Illinois Institute of Technology, Chicago, IL, United States

Automated cerebral microbleed (CMB) detection on ex-vivo MRI is key to enabling MRI-pathology studies in large community-based cohorts where manual CMB annotation is time consuming and prone to error. The aim of this study is to develop a CMB detection algorithm to aid in the quantization and localization of CMBs on ex-vivo T2*-weighted gradient echo MRI in community-based cohorts. A CMB synthesis algorithm is proposed and synthetic CMBs are used to train a neural network for CMB detection. A model trained with both synthetic and real data is shown to outperform models trained on synthetic or real data alone.

Sarah Schlaeger¹, Katharina Drummer¹, Malek El Husseini¹, Florian Kofler^1,2,3, Nico Sollmann^1,4,5, Severin Schramm¹, Claus Zimmer¹, Dimitrios C. Karampinos⁶, Benedikt Wiestler¹, and Jan S. Kirschke^1
1Department of Diagnostic and Interventional Neuroradiology, Klinikum rechts der Isar, Technical University of Munich, Munich, Germany, ²Department of Informatics, Technical University of Munich, Munich, Germany, ³TranslaTUM - Central Insitute for Translational Cancer Research, Technical University of Munich, Munich, Germany, ⁴TUM-NeuroImaging Center, Klinikum rechts der Isar, Technical University of Munich, Munich, Germany, ⁵Department of Diagnostic and Interventional Radiology, University Hospital Ulm, Ulm, Germany, ⁶Department of Diagnostic and Interventional Radiology, Klinikum rechts der Isar, Technical University of Munich, Munich, Germany

Generative Adversarial Networks (GANs) can synthesize missing Magnetic Resonance (MR) contrasts from existing MR data. In spine imaging, sagittal T2-w fat sat (fs) sequences are an important additional MR contrast next to conventional T1-w and T2-w sequences. In this study, the diagnostic performance of a GAN-based, synthetic T2-w fs is evaluated in a multicenter dataset. By comparing the synthetic T2-w fs with its true counterpart regarding ability to detect spinal pathologies not seen on T1-w and non-fs T2-w, diagnostics agreement, and image and fs quality our work shows that a synthetic T2-w fs delivers valuable information on spine pathologies. 

Adrian Schrader^1,2, Nils Bastian Netzer^1,2, Magdalena Görtz³, Constantin Schwab⁴, Markus Hohenfellner³, Heinz-Peter Schlemmer¹, and David Bonekamp^1
1Division of Radiology, German Cancer Research Center, Heidelberg, Germany, ²Heidelberg University Medical School, Heidelberg, Germany, ³Department of Urology, University of Heidelberg Medical Center, Heidelberg, Germany, ⁴Department of Pathology, University of Heidelberg Medical Center, Heidelberg, Germany

For patients with clinical suspicion for significant prostate cancer, the decision to undergo prostate biopsy can be supported by calculating the individual risk profile using demographic and clinical information along with multiparametric MRI assessment. We could show that the prediction performance of an established risk calculator remained stable after substituting manual PI-RADS scores for assessments from a fully automated deep learning system. Combining deep learning and PI-RADS resulted in significant improvements over using only PI-RADS. Complementary information that deep learning models are able to extract enable synergies with radiologists to improve individual risk predictions.

Nils Bastian Netzer^1,2, Adrian Schrader^1,2, Magdalena Görtz³, Constantin Schwab⁴, Markus Hohenfellner³, Heinz-Peter Schlemmer¹, and David Bonekamp^1
1Radiology, German Cancer Research Center, Heidelberg, Germany, ²Heidelberg University Medical School, Heidelberg, Germany, ³Urology, University of Heidelberg Medical Center, Heidelberg, Germany, ⁴Pathology, University of Heidelberg Medical Center, Heidelberg, Germany

The value of dynamic contrast enhanced MRI (DCE) for the diagnosis of prostate cancer is unclear and has not yet been investigated in the context of deep learning. We trained 3D U-Nets to segment prostate cancer on bi-parametric MRI and on DCE images of 761 exams. On a test set of 191 exams,  the bi-parametric baseline achieved a ROC AUC of 0.89, showing a higher specificity that clinical PI-RADS at a sensitivity of 0.9. Additional improvement could be achieved by fusing bpMRI and DCE predictions, resulting in a ROC AUC of 0.9.

Buse Buz-Yalug¹, Ayca Ersen Danyeli^2,3, Cengiz Yakicier^3,4, Necmettin Pamir^3,5, Alp Dincer^3,6, Koray Ozduman^3,7, and Esin Ozturk-Isik^1
1Institute of Biomedical Engineering, Bogazici University, Istanbul, Turkey, ²Department of Medical Pathology, Acibadem Mehmet Ali Aydinlar University, Istanbul, Turkey, ³Center for Neuroradiological Applications and Reseach, Acibadem Mehmet Ali Aydinlar University, Istanbul, Turkey, ⁴Department of Molecular Biology and Genetics, Istanbul, Turkey, ⁵Department of Neurosurgery, Acibadem Mehmet Ali Aydinlar University, Istanbul, Turkey, ⁶Department of Radiology, Acıbadem Mehmet Ali Aydinlar University, Istanbul, Turkey, ⁷Department of Neurosurgery, Acıbadem Mehmet Ali Aydinlar University, Istanbul, Turkey

Isocitrate dehydrogenase (IDH) and telomerase reverse transcriptase promoter (TERTp) mutations highly affect the clinical outcome in gliomas. The aim of this study was to identify IDH and TERTp mutations in glioma patients using machine learning approaches on relative cerebral blood volume (rCBV) maps obtained from dynamic susceptibility contrast MRI (DSC-MRI). The highest classification accuracy was 87.2% (sensitivity = 85.7%, specificity = 88.9%) for the IDH subgroup, 81.8% accuracy (sensitivity = 77.5%, specificity = 86.4%) was obtained for classifying the TERTp subgroup. Additionally, a classification accuracy of 89.6% (sensitivity = 88.3%, specificity = 91.2%) was obtained for identifying the TERTp-only gliomas.

Gaurav Verma¹, Xin Xing², Yael Jacob³, Bradley N Delman⁴, James Murrough³, Ai-Ling Lin⁵, and Priti Balchandani^1
1Biomedical Engineering and Imaging Institute, Icahn School of Medicine at Mount Sinai, New York, NY, United States, ²Computer Science, University of Kentucky, Lexington, KY, United States, ³Psychiatry, Icahn School of Medicine at Mount Sinai, New York, NY, United States, ⁴Diagnostic, Molecular and Interventional Radiology, Icahn School of Medicine at Mount Sinai, New York, NY, United States, ⁵Radiology, University of Missouri, Columbia, MO, United States

Major depression is highly-prevalent disorder with frustratingly-high rates of treatment resistance. Ultrahigh field imaging may provide objective quantitative biomarkers for characterizing depression, generating insight into clinical phenotypes of this heterogeneous disease. Forty-two major depressive disorder patients currently off anti-depressant treatment were recruited for scanning at ultrahigh field, and given batteries of clinical symptom measures. Machine-learning clustering analysis was performed to group patients by clinical symptoms and differences in imaging features observed. A separate analysis was performed in the reverse direction clustering on quantified imaging features and identifying clinical differences between clusters, including differences in ruminative response between the patient clusters.

Eugene Ozhinsky¹, Felix Liu¹, Valentina Pedoia¹, and Sharmila Majumdar^1
1Radiology and Biomedical Imaging, University of California San Francisco, San Francisco, CA, United States

High quality scan prescription that optimally covers the area of interest with scan planes aligned to relevant anatomical structures is crucial for error-free radiologic interpretation. In this study we used images and metadata from previously acquired examinations of lumbar spine to train machine learning-based automated prescription models without the need of any manual annotation or feature engineering. The automated prescription pipeline was integrated with the scanner console software and evaluated in healthy volunteer experiments. This study demonstrates the feasibility of using oriented object detection-based pipelines on the scanner for automated prescription of lumbar spine acquisitions.

Paul E Summers¹, Lars Johannes Isaksson², Matteo Johannes Pepa², Mattia Zaffaroni², Maria Giulia Vincini², Giulia Corrao^2,3, Giovanni Carlo Mazzola^2,3, Marco Rotondi^2,3, Sara Raimondi⁴, Sara Gandini⁴, Stefania Volpe^2,3, Zaharudin Haron⁵, Sarah Alessi¹, Paola Pricolo¹, Francesco Alessandro Mistretta⁶, Stefano Luzzago⁶, Federico Cattani⁷, Gennaro Musi^3,6, Ottavio De Cobelli^3,6, Marta Cremonesi⁸, Roberto Orecchia⁹, Giulia Marvaso^2,3, Barbara Alicja Jereczek-Fossa^2,3, and Giuseppe Petralia^3,10
1Division of Radiology, IEO, European Institute of Oncology IRCCS, Milano, Italy, ²Division of Radiation Oncology, IEO, European Institute of Oncology IRCCS, Milano, Italy, ³Department of Oncology and Hemato-oncology, University of Milan, Milano, Italy, ⁴Department of Experimental Oncology, IEO, European Institute of Oncology IRCCS, Milano, Italy, ⁵Radiology Department, National Cancer Institute, Putrajaya, Malaysia, ⁶Division of Urology, IEO, European Institute of Oncology IRCCS, Milano, Italy, ⁷Unit of Medical Physics, IEO, European Institute of Oncology IRCCS, Milano, Italy, ⁸Radiation Research Unit, IEO, European Institute of Oncology IRCCS, Milano, Italy, ⁹Scientific Directorate, IEO, European Institute of Oncology IRCCS, Milano, Italy, ¹⁰Precision Imaging and Research Unit, IEO, European Institute of Oncology IRCCS, Milano, Italy

A persisting concern is that downstream models of clinical endpoints may depend on whether the contours were drawn by an expert or an AI. Prediction models for surgical margin status, and pathology-based lymph nodes, tumor stage and ISUP grade group were formed using clinical and radiological features along with whole-prostate radiomic features based on manual and AI segmentations of the prostate in 100 patients who proceeded to prostatectomy after multiparametric-MRI. The models based on AI segmented prostates differed from those based on manual segmentation, but with similar if not better performance. Further testing of generalizability of the models is required.

Erin Beate Bjørkeli^1,2, Jonn Terje Geitung^1,2, and Morteza Esmaeili^1,3
1Department of Diagnostic Imaging, Akershus University Hospital, Oslo, Norway, ²Institue of Clinical Medicine, University of Oslo, Oslo, Norway, ³Department of Electrical Engineering and Computer Science, University of Stavanger, Stavanger, Norway

We have developed a MRSI spectra classifying convolutional neural network (CNN), building on the AUTOMAP model for image and spectra reconstruction. The model was trained to discern between non-water-suppressed spectra from healthy subjects and glioblastoma (GBM) patients. The trained CNN was able to classify the spectra correctly and seemed to recognize the healthy spectra based on the NAA-peak and the GBM based on the choline levels and possibly 2HG, indicating an IDH mutation.

Marion Tardieu¹, Lakhdar Khellaf², Maida Cardoso³, Olivia Sgarbura⁴, Pierre-Emmanuel Colombo⁴, Christophe Goze-Bac³, and Stephanie Nougaret^1
1Montpellier Cancer Research Institute (IRCM), INSERM U1194, University of Montpellier, Montpellier, France, ²Department of pathology, Montpellier Cancer Institute (ICM), Montpellier, France, ³BNIF facility, L2C, UMR 5221, CNRS, University of Montpellier, Montpellier, France, ⁴Department of Surgery, Montpellier Cancer Institute (ICM), Montpellier, France

The objective was to probe the associations of high-field MR-images and their derived texture maps (TM) with histopathology in ovarian cancer (OC). Four ovarian tumors were imaged ex-vivo using a 9.4T-MR scanner. Automated MR-derived stroma-tumor segmentation maps were constructed using machine learning and validated against histology. Through TM, we found that areas of tumor cells appeared uniform on MR-images, while areas of stroma appeared heterogeneous.  Using the automated model, MRI predicted stromal proportion with an accuracy from 61.4% to 71.9%. In this hypothesis-generating study, we showed that it is feasible to resolve histologic structures in OC using ex-vivo MR radiomics.

Hannes Schreiter^1,2, Vishal Sukumar^1,2, Lorenz Kapsner¹, Lukas Folle², Sabine Ohlmeyer¹, Frederik Bernd Laun¹, Evelyn Wenkel¹, Michael Uder¹, Andreas Maier², Sebastian Bickelhaupt¹, and Andrzej Liebert^1
1Institute of Radiology, University Hospital Erlangen, Friedrich-Alexander University Erlangen-Nürnberg (FAU), Erlangen, Germany, ²Patter Recognition Lab, Department of Computer Science, Friedrich-Alexander University Erlangen-Nürnberg (FAU), Erlangen, Germany

Magnetic resonance imaging (MRI) examinations of the breast require intravenous administration of gadolinium based contrast agents (GBCA) for comprehensive characterization of the tissue. Novel approaches reducing the need for GBCA might therefore be of value. Here a virtual dynamic contrast enhancement (vDCE) using a U-net architecture is investigated in a cohort of n=540 patients. The vDCE generates T1 subtraction images for five consecutive time points predicting the perfusion maps based on native T1-weighted, T2-weighted, and multi-b-value diffusion weighted acquisitions. A mean structural similarity index (SSIM) value over a test group of 82 patients of 0.848±0.025 was achieved.