Photo: @SaraBarriere

Welcome to the Swedish symposia on deep learning and image analysis!


SSDL/SSBA 2022 goes virtual!

Due to the pandemic, the symposium can not be held on site in Uppsala, but will be a virtual event. Participation will be free-of-charge for all. SSDL/SSBA 2022 thereby marks the continuation of a 39-year-old tradition of gathering researchers in the field of image analysis and pattern recognition who are active in Sweden to share their latest findings and socialize in an annual national forum.

SSDL is an important Swedish forum for leading research groups in industry and academia to meet and discuss the latest trends and developments in deep learning and related areas. SSDL 2022 will feature invited talks by leading researchers in deep learning as well as oral and poster presentations of submitted papers and abstracts.



The SSBA symposium is the premier Swedish event where researchers, industrial professionals and students gather to learn about the recent developments in the areas of image processing, computer vision, pattern recognition and related fields. SSBA 2022 features keynote speakers and oral presentations and posters of submitted papers.


SSDL/SSBA 2022 is organized by the Department of Information Technology at Uppsala University, Centre for Image Analysis, and Swedish Society for Automated Image Analysis.

Important dates

Deadlines
SSBA paper submission deadline Passed
SSDL abstract/paper submission deadline Passed

Registration deadline

Passed



Conference dates
Deep Learning Symposium (SSDL) 14 March 12.00-18:00
Image Analysis Symposium (SSBA) 15 - 16 March

Program






Keynote speakers



Amy Loutfi
  • Amy Loutfi
  • Örebro University

  • Deep Learning in Collaborative Human Robot Interaction Scenarios.
  • I will speak about how the need for robots and other artificial agents to truly understand human instruction increases. This understanding involves the ability for robots and artificial agents to gain semantic meaning that refer to physical objects, actions, and concepts, and connect this to the perceptual information coming from the sensors embedded on the agent. This challenge in part deals with the symbol grounding problem, that is, how to ground symbols into something other than symbols. Further with the advent of advances within representation learning, due to many of the advances within sub-symbolic architectures there has been an opportunity to examine how the grounding can be learned using large (and simulated) datasets. Further, this talk will describe how the semantic representations can also further guide sub-symbolic learning. This talk will highlight the research at Örebro University where semantics are learned, integrated and used for collaborative robot systems that work both for and with humans.



Orcun Göksel
  • Orcun Göksel
  • Uppsala University

  • Deep Learning in Medical Ultrasound Imaging: Simulation and Image Reconstruction
  • Ultrasound is a cost-effective, portable, and real-time imaging modality, involving minimal safety risks for the patient and the sonographer. In this talk, our recent advances in deep learning for ultrasound image analysis will be presented in two different settings. In the first setting, learning image translation from unpaired samples will be demonstrated for the computational simulation of ultrasound imaging, which is essential for training sonographers for complex imaging tasks in a simulated virtual-reality environment. In a second setting, our solutions with deep variational networks in image reconstruction will be demonstrated, in particular for loop-unrolling iterative solutions of inverse problems that occur in ultrasound computed tomography. This is presented for a novel imaging approach for tissue speed-of-sound, which is a promising new modality for breast cancer and muscular degeneration.

    Orcun Goksel is associate professor at the Department of Information Technology at Uppsala University. He leads the research group Computer-assisted Applications in Medicine, as part of Uppsala Medtech Science and Innovation Centre and the Centre for Image Analysis. He received bachelor degrees in computer science and electrical engineering. Following a PhD degree from University of British Columbia, in Vancouver, Canada, he held a Swiss National Science Foundation Professorship at ETH Zurich, Switzerland, before joining Uppsala University. His research interests include machine learning, computer vision, ultrasound imaging in biomedical applications, biomechanical tissue characterization, image reconstruction, image-guided therapy, and patient-specific modelling and simulation. By devising novel imaging and image analysis techniques and developing them for clinical translation, his research efforts push the boundaries of diagnostic and surgical procedures as well as minimally-invasive interventions.



Cynthia Rudin
  • Cynthia Rudin
  • Duke University

  • Interpretable Neural Networks for Computer Vision: Clinical Decisions that are Computer-Aided, not Automated
  • Let us consider a difficult computer vision challenge. Would you want an algorithm to determine whether you should get a biopsy, based on an x-ray? That's usually a decision made by a radiologist, based on years of training. We know that algorithms haven't worked perfectly for a multitude of other computer vision applications, and biopsy decisions are harder than just about any other application of computer vision that we typically consider. The interesting question is whether it is possible that an algorithm could be a true partner to a physician, rather than making the decision on its own. To do this, at the very least, we would need an interpretable neural network that is as accurate as its black box counterparts. In this talk, I will discuss two approaches to interpretable neural networks: (1) case-based reasoning, where parts of images are compared to other parts of prototypical images for each class, and (2) neural disentanglement, using a technique called concept whitening. The case-based reasoning technique is strictly better than saliency maps, and the concept whitening technique provides a strict advantage over the posthoc use of concept vectors. Here are the papers I will discuss:

    This Looks Like That: Deep Learning for Interpretable Image Recognition. NeurIPS spotlight, 2019. https://arxiv.org/abs/1806.10574

    IAIA-BL: A Case-based Interpretable Deep Learning Model for Classification of Mass Lesions in Digital Mammography, 2021. https://arxiv.org/abs/2103.12308

    Concept Whitening for Interpretable Image Recognition. Nature Machine Intelligence, 2020. https://rdcu.be/cbOKj

    Stop Explaining Black Box Machine Learning Models for High Stakes Decisions and use Interpretable Models Instead, Nature Machine Intelligence, 2019. https://rdcu.be/bBCPd

    Interpretable Machine Learning: Fundamental Principles and 10 Grand Challenges, 2021 https://arxiv.org/abs/2103.11251



David J.T. Sumpter
  • David J.T. Sumpter
  • Uppsala University

  • Why mathematical modelling needs to be more creative and unconventional
  • One of the challenges for modellers trying to understand complex biological and social systems is that cultures build up around what are ‘interesting’ research topics. My own academic career (starting 1996) has seen crazes for self-organisation, evolutionary game theory, network science and now machine learning. All of these approaches brought interesting theoretical and applied results, but seen over the longer term, we also notice a recurring problem: Craze X takes on a greater importance than the Science itself. Applications are all re-envisioned in light of Craze X and theoretical debates become about the (often meaningless) details of Craze X. With reference to my own research and popular science writing, I describe how I think we can escape this trap, through a combination of two things: extremely focused applications and extremely unfocussed theoretical investigations. I illustrate the former with my work on modelling football. I illustrate the latter with an example from Artificial Life. I will talk about how wrongly focusing on the middle ground – the development of computational and mathematical theory – can lead to dangerous and unnecessary hype, which can then have negative social impacts.



Marcus Liwicki
  • Marcus Liwicki
  • Luleå University of Technology

  • Historical Document Analysis – history, trends, and challenges
  • I will speak about the development in the field of historical document image analysis during the last two decades (before and after deep learning), the current development, as well as open challenges. Historical document image analysis remains one of the most challenging topics for the document image analysis research community. Historical documents differ from the ordinary documents due to the presence of different artifacts. Issues such as poor conditions of the documents, texture, noise and degradation, large variability of page layout, page skew, random alignment, variety of fonts, presence of embellishments, variations in spacing between characters, words, lines, paragraphs and margins, overlapping object boundaries, superimposition of information layers, etc bring complexity issues in analyzing them. Most methods currently rely on deep learning based methods, including Convolutional Neural Networks and Long Short-Term Memory Networks. As those methods are data-hungry, a recent trend is to generate more training data with various methods. In our VR project, we are dealing with this challenge and working on a systematic approach towards training sets with millions of documents.



Virginie Uhlmann
  • Virginie Uhlmann
  • European Bioinformatics Institute (EMBL-EBI) in Cambridge, UK

  • Deep learning for bioimage analysis: how far we got, and where we are heading
  • Deep learning has transformed the way large and complex microscopy image datasets can be processed, reshaping what is possible in bioimage analysis. In this talk, we will characterize the unique features and challenges of microscopy data, and review how deep learning methods have successfully been adapted from computer vision and medical imaging to efficiently solve recurring bioimage analysis problems. We will also discuss the future of deep learning applied to biology, and explore how state-of-the-art approaches in machine learning have the potential to transform our understanding of biological systems through new analysis and modelling strategies that integrate multimodal inputs in space and time.



Submissions, SSDL (Password protected)

The submissions for the SSDL poster session is listed below. Some of the submissions are also presented as regular talks.

Gryska et al. Updated reporting checklist for deep-learning based automatic brain tumour segmentation on MRI (poster DL1)
Karlsson et al. Classification of breast ultrasound images using different machine learning approaches (poster DL2 + oral)
Kartasalo et al. Artificial Intelligence for Detecting Perineural Invasion in Prostate Biopsies (poster DL3)
Liu et al. A machine learning approach to resolving time dependent ODEs and PDEs with noise (poster DL4)
Poceviciute et al. Uncertainty and reliability of DL for digital pathology applications (poster DL5)
Ji et al. Physical Color Calibration of Digital Pathology Scanners for Deep Learning Based Diagnosis of Prostate Cancer (poster DL6 + oral)
Chatterje et al. DCNN based Oral Cancer Screening using Whole Slide Cytology Images: Effect of Increased Patch Size (poster DL7)
Kahramana et al. Automated Computer-Aided Pulmonary Embolism Detection in CT pulmonary angiography volume images (poster DL8)
Liu et al. Calibrated Energy-based Models for Deep Probabilistic Regression (poster DL9)
Baravdish et al. Learning via nonlinear conjugate gradients and depth-varying neural ODEs (poster DL10)
Mulliqi et al. A robust artificial intelligence approach for histopathological evaluation of prostate biopsies (poster DL11)
Walhagen et al. PAS – an AI model trained exclusively on patient outcome data beats world leading pathologists in prostate cancer risk grading (poster DL12)
Jönemo et al. Evaluation of Augmentation Methods in Classifying Autism Spectrum Disorders from FMRI Data with 3D Convolutional Neural Networks (poster DL13)
Hsu et al. Attention Guided Low-light Image Enhancement with Scene Text Restoration (poster DL14)
Gustafsson et al. Learning Proposals for Practical Energy-Based Regression (oral)
Adewumi et al. Småprat: DialoGPT for Natural Language Generation of Swedish Dialogue by Transfer Learning (oral)

Submissions, SSBA (Password protected)

The submissions for the SSBA poster session is listed below. Some of the submissions are also presented as regular talks.

All SSBA submissions are also available in the SSBA proceedings.

Andersson et al. FLIP: Visualizing and Communicating the Differences Between Images (oral)
Åström et al. A Low cost and High dynamic Time-to-Impact Estimation Sensor Algorithm Based on a Novel Statistical Approach (oral)
Bezek et al. Mean Speed-of-Sound Estimation Using Geometric Disparities (oral)
Breznik et al. Label-free reverse image search of multimodal microscopy images (poster BA7 + oral)
Brynte et al. Adapting the Translation Equivariance of CNNs for 3D Inference Tasks (poster BA6 + oral)
Flood et al. Map Matching with Statistical Deformations Using Minimal Solvers (oral)
Gummeson et al. Minimal solvers for camera pose estimation from cylinders and quadrics (poster BA9)
Hermanova et al. Quantitative characteristics of brain lesions: a data visualization tool (poster BA2 + oral)
Iglesias et al. Radial Distortion Invariant Factorization for Structure from Motion (oral)
Jonnarth et al. Activation-Based Sampling for Pixel- to Image-Level Aggregation in Weakly-Supervised Segmentation (poster BA3 + oral)
Sabel et al. Evaluation of 3d point cloud registration over forests for aerial localisation (oral)
Siddiqui et al. FEXGAN-META: Facial expression generation with meta humans (poster BA4 + oral)
Pielawski et al. GPU-accelerated interactive quality assessment and visualization of large-scale spatial omics and digital pathology data (poster BA1 + oral)
Wetzer et al. Rotationally Equivariant Representation Learning for Multimodal Images (poster BA5 + oral)
Yngesjö et al. 3D Reconstruction from Satellite Imagery Using Deep Learning (poster BA8 + oral)
Öfverstedt et al. Efficient Algorithms for Global Multimodal Image Registration (oral)

Venue

Venue: Online

If you have any questions please contact us on ssbassdl2022@ssba.org.se

Organizers

Industry support

Last Updated Date: 14 March 2022