Abstracts

ABSTRACTS

____________________________________________________________________________________

Session 1 : Morning

____________________________________________________________________________________

Talk 1: Stereo Vision Based Surface Registration for Abdomen Minimally Invasive Surgery

Abstract

To improve the surgery safety and accuracy, computer vision techniques are explored to assist surgeon operating Minimally Invasive Surgery. In this presentation, I will introduce our work including segmentation of the abdomen organs before surgery on CT images, surgery video analysis with a focus on image quality improvement, 3D reconstruction of the surgery scene.

Talk 2 : The Next Generation of AI in Medicine

Abstract

Deep Learning (DL) has emerged as a leading technology in computer science for accomplishing many challenging tasks. This technology shows an outstanding performance in a broad range of computer vision and medical applications. However, this success comes at the cost of collecting and processing a massive amount of data, which are in healthcare often inaccessible due to privacy issues. Federated Learning is a new technology that allows training DL models without sharing the data. Using Federated Learning, DL models at local hospitals share only the trained parameters with a centralized DL model, which is, in return, responsible for updating the local DL models as well. Yet, a couple of well-known challenges in the medical imaging community, e.g., heterogeneity, domain shift, scarify of labeled data and handling multi-modal data, might hinder the utilization of Federated Learning. In this talk, a couple of proposed methods, to tackle the challenges above, will be presented paving the way to researchers to integrate such methods into the privacy-preserved federated learning.

Talk 3 : Residual network based distortion classification and ranking for laparoscopic image quality

Abstract

Laparoscopic images and videos are often affected by different types of distortion like noise, smoke, blur and nonuniform illumination. Automatic detection of these distortions, followed generally by application of appropriate image quality enhancement methods, is critical to avoid errors during surgery. In this context, a crucial step involves an objective assessment of the image quality, which is a two-fold problem requiring both the classification of the distortion type affecting the image and the estimation of the severity level of that distortion. Unlike existing image quality measures which focus mainly on estimating a quality score, we propose in this work to formulate the image quality assessment task as a multi-label classification problem considering both the type as well as the severity level (or rank) of distortions. Here, this problem is then solved by resorting to a deep neural networks based approach. The obtained results on a laparoscopic image dataset show the efficiency of the proposed approach.

Talk 4: New horizons in deep learning-assisted multimodality medical image analysis

Abstract

Positron emission tomography (PET), x-ray computed tomography (CT) and magnetic resonance imaging (MRI) and their combinations (PET/CT and PET/MRI) provide powerful multimodality techniques for in vivo imaging. This talk presents the fundamental principles of multimodality imaging and reviews the major applications of artificial intelligence (AI), in particular deep learning approaches, in multimodality medical image analysis. It will inform the audience about a series of advanced development recently carried out at the PET instrumentation & Neuroimaging Lab of Geneva University Hospital and other active research groups. To this end, the applications of deep learning in five generic fields of multimodality medical imaging, including imaging instrumentation design, image denoising (low-dose imaging), image reconstruction quantification and segmentation, radiation dosimetry and computer-aided diagnosis and outcome prediction are discussed. Deep learning algorithms have been widely utilized in various medical image analysis problems owing to the promising results achieved in image reconstruction, segmentation, regression, denoising (low-dose scanning) and radiomics analysis. This talk reflects the tremendous increase in interest in quantitative molecular imaging using deep learning techniques in the past decade to improve image quality and to obtain quantitatively accurate data from dedicated combined PET/CT and PET/MR systems. The deployment of AI-based methods when exposed to a different test dataset requires ensuring that the developed model has sufficient generalizability. This is an important part of quality control measures prior to implementation in the clinic. Novel deep learning techniques are revolutionizing clinical practice and are now offering unique capabilities to the clinical medical imaging community. Future opportunities and the challenges facing the adoption of deep learning approaches and their role in molecular imaging research are also addressed.

Talk 5 : Numerical observers for the objective quality assessment of medical images.

Abstract

Medical image quality assessment is critical and indispensable for comparing and optimizing medical imaging systems such as acquisition systems, image post-processing systems and visualization systems. Since the ultimate goal of medical images is to aid clinicians in rendering a diagnosis, it is widely accepted that in order to optimize diagnostic decisions image quality should be as good as possible. Additionally, while it has been poorly addressed by natural image quality assessment community, task-based approaches are fundamental and popular in the context of medical imaging. The underlying paradigm is to quantify the quality of a particular image by its effectiveness with respect to its intended task. Natural images quality assessment is usually focused on the perception of impairments, while the quality assessment of medical images normally focus on the radiologists’ diagnostic task performance. Many numerical observers have been proposed in the framework of task-based approach. In this talk, the principal of the task-based approach and the basics of the numerical observers will be presented.

_______________________________________________________________________

Session 2 : afternoon

_______________________________________________________________________

Talk 6 : Computer aided diagnoses for wireless capsule endoscopy

Abstract

The progress in Computer Aided Diagnosis (CADx) of Wireless Capsule Endoscopy (WCE) is thwarted by the lack of data. The inadequacy in richly representative healthy and abnormal conditions results in isolated analyses of pathologies, that can not handle realistic multi-pathology scenarios. In this talk, we discuss the challenges arising from this medical modality and explore the possibilities for learning more efficiently, from limited data and even fewer annotations.

Talk 7 : Learning the COVID-19 Proteomic mechanisms from a graph convolutional network of topological models

Abstract

Many efforts have been recently done to characterize the molecular mechanisms of COVID-19 disease. These efforts resulted in a full structural identification of ACE2 as principal receptor of the SARS-Cov2 spike protein in the cell. However, there are still important open questions related to other proteins involved in the progression of the dis-ease. To this end, we have modelled the plasma proteome of 384 COVID patients. The model calibrated proteins measures at 3 time tags and make also use of the detailed clinical evaluation outcome of each patient after leaving the hospital at day 28. Our analysis is able to discriminate severity of the disease by means of a metric based on available WHO scores of disease progression. Then, we identify by topological vectorization those proteins shifting the most in their expression depending on that severity classification. Finally, the extracted topological invariants respect the protein expression at different times were used as base of a graph convolutional network. This model enabled the dynamical learning of the molecular interactions produced between the identified proteins.

Talk 8: Automatic detection and spatial modelling of Ulcerative Colitis lesions from colonoscopy videos

Abstract

Ulcerative colitis is a lifelong idiopathic disorder of the gastrointestinal tract that is currently increasing in prevalence worldwide. It causes lesions such as bleeding and ulcers that begin in the rectum and continually spread along the colon. The assessment of the severity of the disease is based on the most severe lesions found in the colonoscopy video. In this talk, we present an algorithm for detecting all the lesions found in the colonoscopy video using suitable color spaces. Then, we propose to model the spatial distribution of the lesions using the reaction diffusion equation of type Fisher Kolmogorov-Petrovski-Puskinov (FKPP) in order to estimate the long time behavior of the disease propagation. The presented results are obtained based on a custom database build due to the collaboration between LAGA and the Bichat-Beaujon hospitals (Drs. Éric Ogier Denis and Xavier Tréton).

Talk 9: Towards a comprehensive database to study the impact of image quality on abnormality detection and classification in Wireless Capsule Endoscopy

Abstract

This presentation is part of the work we have been doing for the last few months where we are interested in tools and methods for the classification of abnormalities in endoscopic capsule image sequences of the gastrointestinal tract. It concerns the creation of a database that contains different scenarios and in particular the types of distortions that can affect the performance of deep learning- based classification algorithms. We will outline the challenging problems, show some preliminary results and describe some avenues of research in the field of image quality improvement and their contribution to high-level tasks in the context of medical diagnosis of some typical gastrointestinal cancers.