Lipei Zhang¹, Yiran We², Chao Li^1,2, Stephen John Price², and Carola Bibiane Schönlieb^1
1The Centre of Mathematical Imaging in Healthcare, Department of Applied Mathematics and Theoretical Physics, University of Cambridge, Cambridge, United Kingdom, ²Cambridge Brain Tumor Imaging Laboratory, Division of Neurosurgery, Deoartment of Clinic Neuroscience, Univerisity of Cambridge, Cambridge, United Kingdom

Automatic tumor segmentation on MRI is crucial for patient management of brain tumors. However, previous automatic MRI segmentation methods based on deep learning are limited by bottlenecks of feature extraction. In this paper, we seek to investigate the influence of incorporating prior histology grading on the performance of tumor segmentation. We propose to employ histological grade labels to strengthen the feature extraction and tumor localization in the encoder, which can boost the DSC at exceeding 2% in different U-Net baselines. Our results could support the usefulness of incorporating clinical prior knowledge to improve the segmentation performance without introducing extra parameters.
 

Gian Franco Piredda^1,2,3, Punith B. Venkategowda⁴, Piotr Radojewski^5,6, Tom Hilbert^1,2,3, Arun Joseph^6,7,8, Gabriele Bonanno^6,7,8, Roland Wiest^5,6, Karl Egger⁹, Shan Yang⁹, Jean-Philippe Thiran^2,3, Ricardo A. Corredor-Jerez^1,2,3, Bénédicte Maréchal^1,2,3, and Tobias Kober^1,2,3
1Advanced Clinical Imaging Technology, Siemens Healthcare AG, Lausanne, Switzerland, ²Department of Radiology, Lausanne University Hospital and University of Lausanne, Lausanne, Switzerland, ³LTS5, École Polytechnique Fédérale de Lausanne (EPFL), Lausanne, Switzerland, ⁴Siemens Healthcare Pvt. Ltd., Bangalore, India, ⁵Support Center for Advanced Neuroimaging, Institute for Diagnostic and Interventional Neuroradiology, Inselspital, Bern University, Bern, Switzerland, ⁶Translational Imaging Center, sitem-insel AG, Bern, Switzerland, ⁷Advanced Clinical Imaging Technology, Siemens Healthcare AG, Bern, Switzerland, ⁸Magnetic Resonance Methodology, Institute of Diagnostic and Interventional Neuroradiology, University of Bern, Bern, Switzerland, ⁹Department of Neuroradiology, Medical Center – University of Freiburg, Faculty of Medicine, University of Freiburg, Freiburg, Germany

Large spatial signal variations due to field inhomogeneities complicate the application of automated brain morphometry at 7T. In this work, we propose to use transfer learning to adapt a template-based segmentation algorithm to sub-millimeter ultra-high field applications. More specifically, a convolutional neural network pre-trained on T₁-weighted scans to extract the total intracranial volume (TIV) from MP-RAGE acquisitions was re-trained to retrieve the TIV mask directly from MP2RAGE volumes. The developed method proved to reliably deliver brain tissue masks and volumetry at 7T.

Fadila Zerka¹, Mohammed Sunoqrot¹, Bendik Abrahamsen¹, Alexandros Patsanis¹, Tone Frost Bathen^1,2, and Mattijs Elschot^1,2
1Department of Circulation and Medical Imaging, NTNU, Trondheim, Norway, ²Department of Radiology and Nuclear Medicine, St. Olavs Hospital, Trondheim, Norway

Accessing medical data is highly protected by law and ethics, making data sharing difficult and time-consuming. Distributed learning in its various forms allows learning from medical data without these data ever leaving the medical institutions. In this study, we evaluate the Flower federated learning framework for prostate segmentation on T2-Weighted MRI. The results show that the Federated learning framework performs comparably  to the reference (centralized learning) model.

Yasmina Al Khalil¹, Aymen Ayaz¹, Cristian Lorenz², Jürgen Weese², Josien Pluim¹, and Marcel Breeuwer^1,3
1Biomedical Engineering, Eindhoven University of Technology, Eindhoven, Netherlands, ²Philips Research Laboratories, Hamburg, Germany, ³Philips Healthcare, MR R&D - Clinical Science, Best, Netherlands

Deep learning-based segmentation algorithms largely rely on the availability of extensive, clinically representative data. While collecting such data requires vast resources, image simulation has the potential of generating realistic data reproducing a wide range of scanner sequences or parameters. In this work, we present an MR image simulation pipeline and evaluate its potential for training a deep-learning network for segmenting several brain structures in T1-weighted images acquired from real scanners. We additionally demonstrate how to prevent performance degradation from the lack of tissue texture in simulated images by combining statistical texture analysis and filtering on the evaluation image set.

Fiona Young¹, Patrick Hales^1,2, Laura Mancini^3,4, Sotirios Bisdas^3,4, Caroline Micallef^3,4, Tarek Yousry^3,4, Christopher A. Clark¹, Kristian Aquilina⁵, and Jonathan D. Clayden^1
1UCL Great Ormond Street Institute of Child Health, University College London, London, United Kingdom, ²Department of Radiology, Great Ormond Street Hospital for Children, London, United Kingdom, ³Lysholm Department of Neuroradiology, The National Hospital for Neurology & Neurosurgery, University College London Hospitals NHS Foundation Trust, London, United Kingdom, ⁴Neuroradiological Academic Unit, Department of Brain Repair and Rehabilitation, UCL Queen Square Institute of Neurology, London, United Kingdom, ⁵Department of Neurosurgery, Great Ormond Street Hospital for Children, London, United Kingdom

Tractfinder is an atlas-based approach to segmenting white matter tracts without tractography, in subjects with space-occupying lesions. A shape and orientation prior is combined with fibre orientation distributions in the target data to produce a map of tract location. Results for a given tract can be obtained in a matter of minutes and are comparable to those afforded by clinical tractography. In future, tractfinder could be applied to intraoperative diffusion MRI to aid neuronavigation during brain surgery.

Tianjia Zhu^1,2, Minhui Ouyang¹, Lei Feng³, Kay Sindabizera¹, Jessica Hyland¹, and Hao Huang^1,4
1Department of Radiology, Children's Hospital of Philadelphia, Philadelphia, PA, United States, ²Department of Bioengineering, University of Pennsylvania, Philadelphia, PA, United States, ³Shandong University Cheeloo College of Medicine, Jinan, China, ⁴Department of Radiology, University of Pennsylvania, Philadelphia, PA, United States

On widely used relaxation-based T1-weighted (T1w) and T2-weighted (T2w) MRI images, 0-2-year-old infant brain images exhibit poor gray and white matter contrasts due to dynamic and poor myelination in this stage. Such poor T1w and T2w contrasts hinder accurate segmentation of infant brains based on T1w or T2w images. Instead, diffusion MRI (dMRI)-derived maps offer high contrasts throughout infancy. To leverage rich dMRI contrasts, we established an Infant Brain Segmentation based on Multi-dMRI contrast Attention (BISMA) technique. BISMA fills the gap in accurate deep learning segmentation of infant brain by fusing multi-contrasts of dMRI-derived maps.

Alexander J Daniel¹, Charlotte E Buchanan¹, David M Morris², Hao Li³, Rebecca Noble¹, João Sousa⁴, Steven Sourbron⁴, David L Thomas^5,6,7, Andrew N Priest^3,8, and Susan T Francis^1
1Sir Peter Mansfield Imaging Centre, University of Nottingham, Nottingham, United Kingdom, ²Centre for Cardiovascular Science, University of Edinburgh, Edinburgh, United Kingdom, ³Department of Radiology, University of Cambridge, Cambridge, United Kingdom, ⁴Department of Infection, Immunity and Cardiovascular Disease, University of Sheffield, Sheffield, United Kingdom, ⁵Neuroradiological Academic Unit, UCL Queen Square Institute of Neurology, University College London, London, United Kingdom, ⁶Dementia Research Centre, UCL Queen Square Institute of Neurology, University College London, London, United Kingdom, ⁷Wellcome Centre for Human Neuroimaging, UCL Queen Square Institute of Neurology, University College London, London, United Kingdom, ⁸Department of Radiology, Addenbrooke’s Hospital, Cambridge, United Kingdom

Manual segmentation of the kidneys is very time consuming and reader dependent, this renders measurements of total kidney volume (TKV) in large multi-site studies impractical. Here we use a convolutional neural network (CNN), trained on data from a single MRI vendor, to segment the kidneys of volunteers scanned with a harmonised FSE image protocol on MR scanners from three different vendors (GE, Philips and Siemens). The kidneys were manually segmented by two readers, both of which demonstrated a significant difference in TKV across vendors; no significant difference in TKV was found in the segmentations produced by the CNN.

Aashley S.D. Sardjoe Mishre^1,2, Maaike E. Straat², Borja Martinez-Tellez², Mariëtte R. Boon², Oleh Dzyubachyk^3,4, Andrew G. Webb¹, Patrick C.N. Rensen², and Hermien E. Kan^1
1Department of Radiology, C.J. Gorter Center for High Field MRI, Leiden University Medical Center, Leiden, Netherlands, ²Department of Medicine, Division of Endocrinology and Einthoven Laboratory for Experimental Vascular Medicine, Leiden University Medical Center, Leiden, Netherlands, ³Department of Radiology, Division of Image Processing (LKEB), Leiden University Medical Center, Leiden, Netherlands, ⁴Department of Cell and Chemical Biology, Electron Microscopy section, Leiden University Medical Center, Leiden, Netherlands

Brown adipose tissue (BAT) is considered to be a potential therapeutic target against cardiometabolic diseases. Activated BAT combusts intracellular fatty acids leading to a reduction in fat fraction . Both cold exposure and pharmacological stimuli can activate BAT, but the short-term dynamics of BAT activation are unknown. To assess supraclavicular BAT fat fraction dynamics during cold-exposure, we developed a 1-minute-time-resolution MRI protocol using breath-holds and co-registration to minimize motion-artefacts. We demonstrated the validity and feasibility of our image analysis method, and found an inter-image variability of less than 0.1% fat fraction.

Renyang Gu¹, Michela Antonelli², Pritesh Mehta ³, Ashik Amlani ⁴, Adrian Green⁴, Radhouene Neji ⁵, Sebastien Ourselin², Isabel Dregely², and Vicky Goh^2
1Department of Biomedical Engineering and Imaging Science, King's College London, London, United Kingdom, ²School of Biomedical Engineering & Imaging Sciences, King's College London, London, United Kingdom, ³University College London, London, United Kingdom, ⁴Department of Radiology, Guy’s & St Thomas’ NHS Foundation Trust, London, United Kingdom, ⁵Siemens Healthcare Limited, London, United Kingdom

Multiple myeloma is a heterogeneous bone marrow cancer. Assessment of changes in mean apparent diffusion coefficient (ADC_mean) is helpful for evaluating treatment response but has not been feasible for the whole skeleton due to the time-consuming nature of manual segmentation. Whole-skeleton and per-station ADC_mean were quantified from whole-body MRI using automated segmentation by an uncertainty-aware nnU-Net in 30 patients with plasma cell disorders and compared against the manual segmentation  by radiologists. No differences were observed in whole-skeleton or per-station ADC_mean when using the automatic and manual segmentations. Further investigation is required in a larger dataset, but initial results are promising.

Subin Erattakulangara¹, Karthika Kelat¹, and Sajan Goud Lingala^1,2
1Roy J Carver Department of Biomedical Engineering, University of Iowa, Iowa city, IA, United States, ²Department of Radiology, University of Iowa, Iowa city, IA, United States

We develop a protocol adaptive Stacked transfer learning (STL) U-NET for soft tissue segmentation in dynamic speech MRI. Our approach leverages knowledge from large open-source datasets, and only needs to be trained on small number of protocol specific images (of the order of 20 images). We demonstrate the utility of STL U-NET in efficiently segmenting soft-tissue articulators from three different protocols with different field strengths, vendors, acquisition, reconstruction. Using the DICE similarity metric, we demonstrate segmentation accuracies with our approach to be at the level of manual segmentation.

Nora Vogt¹, Zobair Arya¹, Luis Núñez¹, John Connell¹, and Paul Aljabar^1
1Perspectum, Oxford, United Kingdom

Segmentation of lesions within the liver is pivotal for surveillance, diagnosis and treatment planning, and automated or semi-automated approaches can aid clinical workflows. Many patients that are under surveillance may have had previous surgery, meaning their scans will contain post-surgical features. We investigate the use of a deep learning segmentation model for non-contrast MRI that can distinguish between lesions, surgical clips and resection-induced fluid-filling regions. We report mean dice scores of 0.55, 0.72 and 0.88 for these classes respectively, demonstrating the potential of this model for semi-automated workflows.