Eric Oermann

We’ve seen a decade of advances in machine learning and deep learning with a surge of interest in applying these advances to biomedical topics. What are these advances exactly? How are people looking to apply them to biomedical problems and are we finding any answers or only raising more questions? In short, “are we there yet”? I’ll discuss where we are, where we are going, and how I think that we’ll get “there” as Radiology continues to evolve in the coming years.

Tolga Cukur

MRI offers an unrivaled opportunity to noninvasively examine the structure and function of the human brain. Yet, MRI exams are hindered by limitations on quality and diversity of acquired images due to scan time considerations. Classical approaches to acquisition and processing of imaging data often fail to address these limitations. In this talk, the potential role of machine learning in surpassing these fundamental barriers will be discussed. Novel deep learning techniques for image reconstruction, image synthesis and sampling design will be showcased. State-of-the-art results from these techniques indicate a bright future for machine learning in rapid, high-quality and high-sensitivity neuroimaging.  

Ona Wu

The use of artificial intelligence and machine learning for stroke research and clinical applications are increasing exponentially every year. Applications in stroke research focus on the extraction of phenotypes that can help in diagnosis, prognosis or management of stroke patients. These applications tend to fall into two major categories – classification or segmentation. We will briefly review some of the more frequently used applications in the domains of acute ischemic stroke and hemorrhage detection. We will also discuss some of the potential pitfalls and ethical questions that may arise.

Spyridon Bakas

This talk introduces computational analytics towards speeding up scientific discovery and personalized diagnostics. It then focuses on two areas; a) radiogenomics, b) data-private collaborations. For radiogenomics, a hypothesis-driven study of the first imaging signature of EGFRvIII in glioblastoma, is presented, followed by data-driven studies on EGFRvIII and other EGFR mutations generating hypothesis for deeper investigations. Then, the first federated learning simulation in healthcare is discussed, followed by the first large-scale real-world federation in medicine, towards developing an AI model to detect glioblastoma boundaries based on 10,000 patients scans from >55 international collaborating sites, without sharing any patient data.

Nicholas J Luciw^1,2, William Cameron², Andrew D Robertson³, Sarah Atwi², and Bradley J MacIntosh^1,2
1Department of Medical Biophysics, University of Toronto, Toronto, ON, Canada, ²Sunnybrook Research Institute, Toronto, ON, Canada, ³Department of Kinesiology, University of Waterloo, Waterloo, ON, Canada

Blood-oxygenation-level-dependent (BOLD) MRI contains both neuronally-mediated and physiological contrast, such as cardiac-related hemodynamic pulsatility. Without a coincident cardiac trace, however, few options exist to assess voxel-specific hemodynamic pulsatility. We investigated the feasibility of training a convolutional neural network to generate accurate cardiac-related pulsatility maps without cardiac trace recordings. Using features derived from the BOLD signal, the network produced pulsatility estimates that were significantly associated with ground truth. This automated method enables investigation of cerebrovascular conditions through the vascular contributions to BOLD data, specifically when cardiac trace recordings are unavailable.

Antonio Maria Chiarelli¹, Carlo Sestieri¹, Daniele Mascali¹, Richard Geoffrey Wise¹, and Massimo Caulo^1
1Department of Neuroscience, Imaging and Clinical Sciences, University G. D'Annunzio of Chieti Pescara, Chieti Scalo, Italy

We used Machine Learning (ML) to infer gestational age (GA) at birth, and hence, as a metric of prematurity extent, assess its effect, in 88 premature infants using T2*-w BOLD resting-state connectivity and activity, and T1-w volume in 90 brain regions. ML was able to infer GA at birth. Analysis of the spatial distribution of effects indicated that volumetric alterations, in common with BOLD activity, are partially localized to subcortical structures, but are associated with widespread alterations of connectivity. Our results suggest a potential role for ML in early prediction of neurodevelopmental outcome based on BOLD and anatomical MRI metrics.

Bryan Clifford¹, John Conklin², Susie Huang², Thorsten Feiweier³, Zahra Hosseini⁴, Augusto Lio M. Goncalves Filho², Azadeh Tabari², Serdest Demir², Wei-Ching Lo¹, Maria Gabriela Figueiro Longo², Michael Lev², Pam Schaefer², Otto Rapalino², Kawin Setsompop^5,6, Berkin Bilgic⁷, and Stephen Cauley^7
1Siemens Medical Solutions USA, Inc., Boston, MA, United States, ²Department Radiology, Massachusetts General Hospital, Boston, MA, United States, ³Siemens Healthcare GmbH, Erlangen, Germany, ⁴Siemens Medical Solutions USA, Inc., Atlanta, GA, United States, ⁵Department Radiology, Stanford University, Stanford, CA, United States, ⁶Department of Electrical Engineering, Stanford University, Stanford, CA, United States, ⁷Department Radiology, A. A. Martinos Center for Biomedical Imaging, Massachusetts General Hospital, Harvard Medical School, Boston, MA, United States

This work integrates a novel machine learning-based reconstruction and optimized magnetization transfer preparation modules with a multi-shot echo-planar imaging acquisition to provide comprehensive whole-brain imaging in two minutes. Neuroradiologist evaluation indicated that the proposed method can produce T2, T2*, T1, FLAIR, and DWI images with SNR, tissue contrast, and lesion conspicuity similar to that of a 10-minute turbo spin echo-based exam. To accommodate a wide range of radiologist preferences and/or hardware configurations without the need for additional training, the proposed method provides a tunable parameter for controlling the level of denoising.

Yohan Jun*¹, Yae Won Park*², Yangho Lee¹, Kyunghwa Han², Chansik An³, Seung-Koo Lee², Sung Soo Ahn**², and Dosik Hwang**^1
1Electrical and Electronic Engineering, Yonsei University, Seoul, Korea, Republic of, ²Department of Radiology and Research Institute of Radiological Science and Center for Clinical Imaging Data Science, Yonsei University College of Medicine, Seoul, Korea, Republic of, ³Research and Analysis Team, National Health Insurance Service Ilsan Hospital, Goyang, Korea, Republic of

For the detection of brain metastases, either contrast-enhanced 3D gradient echo (GRE) or spin echo (SE) imaging with black-blood (BB) imaging techniques are commonly used. The objective of this study was to evaluate whether a deep learning (DL) model using both 3D BB imaging and 3D GRE imaging may improve the detection and segmentation performance of brain metastases compared to that using only 3D GRE imaging. We demonstrated that the combined 3D BB and 3D GRE DL model can improve the detection and segmentation performance of brain metastases, especially in detecting small metastases.

Samuel Bobholz¹, Allison Lowman², Michael Brehler², Savannah Duenweg¹, Fitzgerald Kyereme², Elizabeth Cochran³, Jennifer Connelly⁴, Wade Mueller⁵, Mohit Agarwal², Darren O'Neill², Anjishnu Banerjee⁶, and Peter LaViolette^2,7
1Biophysics, Medical College of Wisconsin, Milwaukee, WI, United States, ²Radiology, Medical College of Wisconsin, Milwaukee, WI, United States, ³Pathology, Medical College of Wisconsin, Milwaukee, WI, United States, ⁴Neurology, Medical College of Wisconsin, Milwaukee, WI, United States, ⁵Neurosurgery, Medical College of Wisconsin, Milwaukee, WI, United States, ⁶Biostatistics, Medical College of Wisconsin, Milwaukee, WI, United States, ⁷Biomedical Engineering, Medical College of Wisconsin, Milwaukee, WI, United States

This study used autopsy tissue in order to develop radio-pathomic models for histopathological features of brain cancer.  These models used T1, T1C, FLAIR, and ADC images from 45 patients as input into bagged regression ensembles for cellularity, cytoplasm, and extracellular fluid, using the aligned autopsy tissue samples as ground truth. These models were able to accurately predict these features and were able to find tumor signatures, such as hypercellularity beyond the traditional contrast-enhancing and FLAIR hyperintense regions. These radio-pathomic maps provide new insights into non-invasive signatures of tumor pathology in the post-treatment state and beyond the contrast enhancing region.

Priyanka Tupe Waghmare¹, Piyush Malpure², Manali Jadhav², Abhilasha Indoria³, Richa Singh Chauhan⁴, Subhas Konar⁵, Vani Santosh³, Jitender Saini⁶, and Madhura Ingalhalikar^7
1E &TC, Symbiosis Institute of Technology, Pune, India, ²Symbiosis Center for Medical Image Analysis, Pune, India, ³National Institute of Mental Health and Neurosciences, Bangalore, India, ⁴National Institute of Mental Health & Neurosciences, Pune, India, ⁵National Institute of Mental Health & Neurosciences, Bangalore, India, ⁶Department of Neuroimaging & Interventional Radiology, National Institute of Mental Health & Neurosciences, Bangalore, India, ⁷Symbiosis Center for Medical Image Analysis and Symbiosis Institute of Technology, Pune, India

In midline gliomas, patients with H3K27M mutation have poor prognosis and shorter median survival. Moreover, since these tumors are located in deep locations biopsy can be challenging with substantial risk of morbidity. Our work proposes a non-invasive deep learning-based technique on pre-operative multi-modal MRI to detect the H3K27M mutation. Results demonstrate a testing accuracy of 69.76% on 51 patients. Furthermore, the class activation maps illustrate the regions that support the classification. Overall, our preliminary results provide a testimony that multimodal MRI can support identifying H3K27M mutation and with further larger studies can be translated to clinical workflow.