Publications scientifiques

Mask uncertainty regularization to improve machine learning based medical image segmentation

Vladimir Groza [1], Benoit Huet [1], Nozha Boujemaa [1] - Affiliations : [1] Median Technologies, Valbonne, France

Poster presented at the IEEE International Symposium on Biomedical Imaging (ISBI) | April 3-7, 2020 | Virtual conference.

Segmentation of the different body structures on CT and MRI scans remains a challenging problem that requires accurate ground truth (GT) segmentation. One of the important aspects is the lack of reliability caused by radiologists annotation disagreement coupled with insufficient quality of the medical images. An independent multiple annotation is needed to overcome frequent disagreements between radiologists decisions mostly on the organs border, which is always blurred and affects on the providing an accurate segmentaKon even on contrasted CT/MRI images. Augmentation techniques aim at increasing model generalization by manipulating inputs in order to enrich set of the variability of images in regards to their GT. We introduce a method to regularize the models learning process by augmenting the information only in the associated GT masks.

Pneumothorax segmentation with effective conditioned post-processing in the chest X-Ray

Vladimir Groza [1], Artur Kuzin [2] - Affiliations: [1] MedianTechnologies, Valbonne, France. [2] X5 Retail Group, Russia. Corresponding author e-mail: vladimir.groza@mediantechnologies.com

Poster presented at the International IEEE conference on Biomedical Imaging | April 3-7, 2020 | Virtual conference.
The pneumothorax can be caused by a blunt chest injury, damage from underlying lung disease or it may occur for no obvious reason at all. This is one of the complex problems for the experts manual detection, which can be solved automaHcally and simplify the clinical workflow. In several situations, lung collapse can turn out as serious threat to life.Proposed method presents new segmentaHon pipeline for the chest X-ray images with the multi step conditioned postprocessing.This approach leads to the significant improvement compare with any « baseline » by the reduction of the totally missed and false positive detections of the pneumothorax collapse regions.

Automatic Fibrosis High Risk Prediction using Computed Tomography Imaging

Elton Rexhepaj [1], Corinne Ramos [1], Nozha Boujemaa [1], Benoit Huet [1] -Affiliations: [1] Median Technologies, Valbonne, France, in col. with: Olivier Lucidarme. Corresponding author e-mail: elton.rexhepaj@mediantechnologies.com

Poster presented at AASLD: The liver meeting 2019 conference | Nov 8-12, 2019 |Boston (USA).

The diagnosis of hepatic fibrosis within liver disease is important for prognosis, stratification for treatment and monitoring of treatment. Although liver biopsy is considered the gold standard for staging fibrosis, it has its limitations due to its invasive nature, sampling error and inter-observer variability. Previous studies have shown that computer tomography (CT) perfusion imaging and splenic radiomics can accurately assess and grade liver fibrosis.

Classification of hepatic tissue in CT: Does IV contrast really matter?

Hubert Beaumont [1] -Affiliations: [1] Median Technologies, Valbonne, France, in col. with: Antoine Iannessi, Fanny Orlhac, Jean-Michel Cucchi. Corresponding author e-mail: hubert.beaumont@mediantechnologies.com

Poster presented at AASLD: The liver meeting 2019 conference | Nov 8-12, 2019 |Boston (USA).

The characterization of biological tissues by means of medical imaging continues to spark a lot of attention, and numerous studies have addressed a broad spectrum of clinical applications while involving most of the imaging modalities. Contrast enhancement products were originally designed for improving human perception, but it is not certain computer assisted analysis still needs it. So far, the impact of IV contrast agents on radiomics-based detection is not fully understood.

The Poincaré-Boltzmann Machine: passing the information between disciplines

Pierre Baudot, PhD [1], Mathieu Bernardi, MscI[1] -Affiliations: [1] Median Technologies, Valbonne, France, in col. with Daniel Bennequin, Monica Tapia, Jean-Marc Goaillard. Corresponding author e-mail: pierre.baudot@mediantechnologies.com

Poster presented at the Geometric Sciences of Information 2019 (GSI2019) conference | Aug 27-29, 2019 |Toulouse (France). Information cohomology is a branch of topological data analysis that allows us to quatify directly statistical dependencies and independences in a given dataset. Theorems establish Shannon entropy as a first cohomology class and mutual information as coboundaries on finite probability space endowed with a random variable chain complex structure. We present some simplicial subcase applications to supervised and unsupervised learning in different contexts: transcriptomic and digits and medical CT image classification.