Registered Data

[00479] Advances in clinically-driven AI image reconstruction and processing

  • Session Time & Room : 1E (Aug.21, 17:40-19:20) @E818
  • Type : Proposal of Minisymposium
  • Abstract : The application of Artificial Intelligence methods, particularly in medical imaging, attracts huge interest from the mathematics and computer science community. Often, AI-based methods provide solutions surpassing image quality metrics compared to traditional methods. However, it's not always obvious if improved image quality metrics necessarily translate to improved clinically-relevant questions such as diagnosis and prognosis. This minisymposium focuses on applied AI methodologies that are clinically-oriented and incorporate direct measures and uses of clinical metrics in the learning or evaluation sections. We will cover topics including advances in model-based learned reconstruction, AI-based inverse problems, regularisation, generative models and their clinical applications.
  • Organizer(s) : Ander Biguri, Lorena Escudero
  • Classification : 68Txx, 92Cxx
  • Minisymposium Program :
    • 00479 (1/1) : 1E @E818 [Chair: Ander Biguri & Lorena Escudero]
      • [05254] The impact of model-based ML driven CT reconstruction on tumor segmentation and clinical diagnosis
        • Format : Talk at Waseda University
        • Author(s) :
          • Ander Biguri (University of Cambridge)
          • Carola-Bibiane Schönlieb (University of Cambridge)
          • Lorena Escudero (University of Cambridge)
        • Abstract : Machine learning has recently found success in tomographic image reconstruction, particularly CT. This work explores its impact on image quality, but most importantly, how (or if) that image quality translates to clinically relevant parameters, in this case radiomics. The evaluation of the reconstruction quality thus is moved from image quality metrics like SSIM or PSNR to clinically relevant metrics like predictive power of radiomics based tumour analysis.
      • [03919] Bringing research advances in imaging sciences into the clinic
        • Format : Talk at Waseda University
        • Author(s) :
          • Ozan Öktem (KTH Royal Institute of Technology)
          • Lorena Escudero (University of Cambridge)
          • Thomas Buddenkotte (Jung Diagnostics GmbH)
          • Cathal McCague (University of Cambridge)
          • Carola-Bibiane Schönlieb (University of Cambridge)
          • Evis Sala (University of Cambridge)
        • Abstract : The talk will survey both scientific and practical challenges associated with making state-of-the-art deep learning methods available to clinicians. It will showcase these challenges in an ambitious setting where one seeks to use an end-to-end deep learning based approach for joint reconstruction and downstream post-processing task, the latter being specific for a clinical use case.
      • [02867] Spectral Normalisation of Depthwise Separable Convolutions for Medical Applications
        • Format : Talk at Waseda University
        • Author(s) :
          • Christina Runkel (University of Cambridge)
          • Christian Etmann (University of Cambridge)
          • Michael Moeller (University of Siegen)
          • Carola-Bibiane Schönlieb (University of Cambridge)
        • Abstract : An increasing number of models require the control of the spectral norm of convolutional layers of a neural network. While there is an abundance of methods for estimating and enforcing upper bounds on those during training, they are typically costly in either memory or time. In this talk, we introduce a very simple method for spectral normalisation of depthwise separable convolutions, which introduces negligible computational and memory overhead - allowing to control spectral norms for practical relevant applications like medical imaging.
      • [05266] Mice PET/CT Dataset Augmentation using a 3D Single Image GAN
        • Format : Online Talk on Zoom
        • Author(s) :
          • Jeremy Kim (Stanford University)
          • Jonathan Fisher (Stanford University)
          • Craig Levin (Stanford University)
        • Abstract : In this study, we applied GANs for 3D mice PET/CT data augmentation; these synthetic mice will be used in deep learning (DL) based preclinical research. The lack of available datasets makes it difficult for researchers to apply DL to solve tasks that involve small-animal datasets, such as emission-based attenuation correction for small-animals PET/MR. We applied the Single-Image GAN (SinGAN) Framework to generate multiple realistic synthetic PET/CT scans of mice from a limited number of examples.