Powered by
05:52 CET
ESR/EIBALL - Imaging biomarkers and their combinations in the era of artificial intelligence
Hybrid Imaging Artificial Intelligence and Machine Learning Imaging Informatics Research
Wednesday, February 27, 08:30 - 10:00
Room: M 1
Type of session: Joint Session of the ESR and EIBALL
Topic: Hybrid Imaging, Artificial Intelligence and Machine Learning, Imaging Informatics, Research
Moderators: O. Clément (Paris/FR), N. M. deSouza (Sutton/GB)

A-0056
08:30
Chairpersons' introduction (part 1)
O. Clément; Paris/FR
Learning Objectives

1. To introduce the potential of AI to accelerate the introduction of imaging biomarkers.
2. To introduce the challenges of managing biomarkers with AI systems.
3. To introduce the speakers.

Abstract

Imaging biomarkers have raised a major interest over the last ten years by a radiologist to quantify and monitor diseases. The rapid spread of artificial intelligence in our discipline might trigger a new era where complex information will be more easily integrated. This session focuses on techniques, examples, opportunities and pitfalls in the use if AI for biomarkers development.

A-0057
08:33
Chairpersons' introduction (part 2)
N. M. deSouza; Sutton/GB
Learning Objectives

1. To introduce the potential of AI to accelerate the introduction of imaging biomarkers.
2. To introduce the challenges of managing biomarkers with AI systems.
3. To introduce the speakers.

Abstract

This session will focus on the use of artificial intelligence for selecting and managing imaging biomarkers. Clarification around the use of terms such as ʽartificial intelligenceʼ and ʽbiomarkersʼ will be discussed. The strengths and limitations of the technologies used will be considered. The development of researcher-driven science and technology networks through the EU COST (Co-operation in Science and Technology) action initiative will be showcased using renal biomarkers as an exemplar. Finally, the selection of combinations of biomarkers from hybrid imaging technologies using AI will be addressed.

A-0058
08:35
Building and discovering biomarkers with AI
B. Rance; Paris/FR
Learning Objectives

1. To learn about available supervised vs unsupervised machine learning techniques.
2. To learn about deep-learning methods to discover biomarkers.
3. To understand the strength, but also limits and pitfalls, of machine learning methods.

Abstract

In recent years, Deep Neural Networks (DNN) have achieved unprecedented performances in many domains, especially with the analysis of images. Major results have been announced for several applications, including skin lesions, pneumonia, pathology and so forth. Several algorithms have even been approved by regulatory agencies, e.g. for the diagnosis of diabetic retinopathy in specific circumstances. It now makes no doubt that artificial intelligence algorithms will be part of the medical experts' toolboxes. In this presentation, we will explore the basic principles behind artificial intelligence and neural networks: supervised and unsupervised algorithms, neurons, activation functions, and the overall architectures of networks. We will discuss more specific classes of DNN used today for the exploration of images, namely the convolutional network and the autoencoder-denoiser, and how they can be used to identify new biomarkers. We will emphasise the crucial role played by expert annotations on images, and explore how experts annotations are used to build models. Finally, we will discuss the implications of the use of deep neural networks in medicine and radiology, and briefly explore new risks (for example adversarial attacks) linked with the use of such technologies.

A-0059
09:00
COST action initiatives as a platform for image biomarker selection
A. Caroli; Ranica/IT
Learning Objectives

1. To learn about how COST actions work and the outputs generated.
2. To appreciate examples of a successful COST action for selecting imaging biomarkers.
3. To understand difficulties in selecting biomarkers for specific indications.

Abstract

COST actions are EU-funded, open, and growing pan-European networking tools aimed at building, bridging and expanding interdisciplinary research communities. A successful example is the COST action PARENCHYMA (Magnetic Resonance Imaging Biomarkers for Chronic Kidney Disease - www.renalmri.org), which coordinates the research of leading European groups working on renal MRI. The rising prevalence of Chronic Kidney Disease (CKD) poses a major public health challenge, and an alarming number of negative clinical trials on CKD progression point out the urgent need for better biomarkers to identify patients that are at risk of progression or are likely to respond to candidate therapeutics. MRI biomarkers have shown a high potential to help fill this gap, as they are non-invasive and sensitive to CKD pathophysiology. Based on this rationale, PARENCHYMA is trying to unlock renal MRI biomarker potential by improving their standardisation and availability, and by generating strong multicentre clinical evidence of biological validity and clinical utility. As different imaging biomarkers provide complementary information on pathophysiology, they need to be combined to reach the highest sensitivity. Artificial intelligence could help to identify best disease-specific combinations.

A-0060
09:25
Role of AI in the introduction of imaging biomarkers: accelerator or obstacle?
J. C. Waterton; Manchester/GB
Learning Objectives

1. To learn how AI can improve imaging biomarkers from multiple imaging modalities.
2. To appreciate the challenges "big data" pose to regulators, whether from imaging biomarkers or from the more conventional biospecimen (genomic, proteomic) biomarkers.
3. To understand how best to utilise AI for imaging biomarkers and to avoid the pitfalls.

Abstract

According to the FDA/NIH BEST resource, a biomarker is a "defined characteristic that is measured as an indicator of normal biological processes, pathogenic processes, or responses to exposure or intervention, including therapeutic interventions; ... radiographic characteristics are types of biomarkers". Imaging biomarkers include scores from scoring systems, such as objective tumour response; extensive variables, such as LVEF; and intensive variables, such as the CT Hounsfield Unit. Biomarkers are essential for evidence-based based medicine, for regulatory approvals, and for prescribing information. However biomarkers must be reproducible over time and space: the measured imaging biomarker value must not drift when measured in a different clinic, or when better scanners are introduced. Radiologic examinations often provide imaging biomarkers which guide patient care. However much of the information in single images, or from multi-modal imaging examinations, is not captured in currently available imaging biomarkers. In the hands of an expert radiologist, this additional information further improves patient care: AI offers the hope of creating new biomarkers by quantifying this additional information. AI-derived imaging biomarkers are, however, not exempt from the need to follow established scientific and regulatory validation pathways and roadmaps. They should show which aspect of the underlying pathology is captured by the new biomarker and must demonstrate reproducibility over time and space (including strategies to maintain validity with future scanners yet to be designed). Otherwise, AI-derived imaging biomarkers will not translate and will remain academic curiosities.

09:50
Panel discussion: What infrastructure do we need to exploit AI for selecting, validating and managing imaging biomarkers?
This website uses cookies. Learn more