Organizers: Moo. K. Chung
Abstract: The main aim of the session is to increase the awareness of the Wasserstein distance and optimal transports in medical imaging to the ISBI community. The Wassersstein distance or Kantorovich-Rubinstein metric is a metric defined between two probability distributions. The distance can be viewed as the optimal transport restricted to probability distributions. Numerous studies have demonstrated superior performance over more traditional Euclidean or geometric distances. The method can work particularly well when it is difficult to establish the direct distance between data. The session brings an opportunity to showcase its powerful uses in various applications including hyperbolic embedding, Schrodinger bridges, deep learning and topological data analysis. It is hoped that this session will provide a forum for constructive discussions on Wasserstein distance related methods. The proposed session consists of lectures given by 4 invited researchers.
Organizers: Chulhong Kim
Abstract: High-resolution volumetric optical imaging modalities are growing in their importance for biomedical imaging. However, due to strong light scattering, the penetration depth is limited to ~ 1 mm in biological tissues. Photoacoustic imaging, an emerging hybrid modality that can provide strong optical absorption contrasts, has overcome the fundamental depth limitation of optical imaging by maintaining excellent spatio-temporal resolution representative of ultrasound imaging. The resolution and the maximum imaging depth are scalable with ultrasonic frequency within the reach of diffuse photons. The imaging depth can be up to a few centimeters. Furthermore, photoacoustic imaging can noninvasively deliver anatomical, functional, and molecular information from living tissues. For highly sensitive molecular photoacoustic imaging, exogenous contrast agents with biomarkers are commonly utilized. Thanks to sharing the same signal detection mechanism with conventional ultrasound imaging, photoacoustic imaging can be easily adapted with the existing ultrasound imaging systems. Thus, clinical translation and commercialization should be relatively easy.
In this Special Session, the following topics will be discussed; (1) recent progress on photoacoustic/ultrasound imaging, (2) advanced image processing including deep learning, (3) potential and/or ongoing clinical translation, and (4) industrial perspectives of photoacoustic/ultrasound imaging and challenges for commercialization.
Organizers: Camille Maumet
Abstract: 10 years ago, a series of publications pointed to the difficulty of reproducing scientific findings. This reproducibility crisis was a wake-up call for scientific communities to rethink how we practice and communicate research, and an important driver towards greater transparency and robust results. Ever since, biomedical imaging undertook various efforts to overcome reproducibility issues: From increasing sample sizes for higher statistical power, to data sharing and increased collaborations to acquire such samples, and promoting detailed reporting practices and code sharing to ease computational reproducibility.
But where are we standing with respect to reproducible biomedical imaging now? We discuss recent advances and open questions, and focus on how the conversation has moved beyond efforts to reduce false positive findings to broader questions of generalizability and fairness. How does a finding observed in a given group apply to the population at large? How does a finding obtained with one analysis vary when computed using another tool? How does a finding observed in a given group apply to subgroups of that population, in particular to less represented subgroups? How can open science help with the complex questions of building fair algorithms and fairness in who participates in the process of science?
Organizers: Eliana Vasquez Osorio and Jamie McClelland
Abstract: Images are essential in radiotherapy. They are used in every step of the patient's pathway, including at diagnostic, treatment planning and delivery, and during patient follow-up. Each step has challenges and opportunities that could benefit from clever solutions being developed by researchers working in image computing. However, many of the clever solutions proposed by image computing researchers do not fully address the most important or challenging clinical problems and achieve limited real-world impact.
In this special session we will have four speakers who work in the field of radiotherapy, bringing the domain knowledge to identify the real problems, challenges, and opportunities where image computing solutions can impact millions of patients. The talks will explore the current clinical workflow, the role of image computing in emerging radiotherapy technology, challenges at translating ‘solved' problems to clinical practice (namely segmentation) and end with unsolved challenges and promising solutions.
We will introduce and highlight current problems aiming at enticing the interest of the ISBI community to further build bridges between these fields. If you are a researcher with an interest in solving real world problems, do not miss this session!
Organizers: Pamela Guevara and Jean-François Mangin
Organizers: Dimitrios I. Fotiadis and Karim Lekadir
Organizers: Claire Cury and Julie Coloigner
Abstract: Combining different neuroimaging modalities, such as EEG, fMRI, MEG or diffusion imaging, could expand the knowledge of our brain and exhibit robust biomarkers, more sensitive to pathophysiological changes. Recent research trends focus on integrating functional and structural modalities to enhance high-resolution spatiotemporal neuroimaging information. However, the integration of these modalities remains a real challenge in data fusion modelling of information from different sources (signals/images) and different dynamics. More specifically, the multi-modal setting and information processing, especially when recording EEG during fMRI, implies overcoming technical and methodological aspects. Recently, new procedures have been developed to allow real-time analysis for bimodal neurofeedback studies. This session highlights the recent advances in bimodal data fusion towards better feature extraction for more specific and robust brain functional understanding. The speakers will present their recent methodological development using different combinations of modalities and targeted applications, such as brain rehabilitation via neurofeedback, introducing this neuroscience technique into the ISBI community.
Organizers: Gang Li
Abstract: Deep learning has achieved overwhelming success in learning effective features for 2D/3D images in the Euclidean space. However, many medical imaging data are intrinsically represented as graphs or manifolds in the non-Euclidean space, e.g., brain cortical surfaces and structural/functional networks. Conventional deep learning techniques, especially convolutional neural networks (CNNs), which are inherently limited to 2D/3D grid-structured data, are thus not suited for handling these non-Euclidean medical imaging data. Therefore, many dedicated geometric deep learning (GDL) methods are proposed for learning more effective representations of non-Euclidean data for solving various challenging problems in medical imaging. This session will introduce the development and applications of some advanced GDL methods in medical image analysis, e.g., parcellation, registration, prediction, etc. We will also give a brief discussion on current challenges and future directions of GDL in medical imaging, hoping to arouse the attention of the audience and interest from the ISBI community, thus inspiring more and deeper research achievements in this field.