Back to overview

The future of radiomics: understanding, Validating and Increasing the Specificity of Imaging Biomarkers for personaLizEd medicine (VISIBLE)

English title The future of radiomics: understanding, Validating and Increasing the Specificity of Imaging Biomarkers for personaLizEd medicine (VISIBLE)
Applicant Depeursinge Adrien
Number 179069
Funding scheme Project funding (Div. I-III)
Research institution Laboratoire d'imagerie biomédicale EPFL - STI - IMT - LIB
Institution of higher education University of Applied Sciences and Arts Western Switzerland - HES-SO
Main discipline Information Technology
Start/End 01.09.2018 - 31.08.2022
Approved amount 491'864.00
Show all

All Disciplines (2)

Information Technology
Biomedical Engineering

Keywords (13)

web platforms; radiology; nuclear medicine; molecular imaging; deep learning; oncology; convolutional neural networks; radiomics; interpretability; medical image analysis; open-access; texture analysis; personalized medicine

Lay Summary (French)

De récentes études dans le domaine de recherche appelé « radiomics » ont démontré que des Biomarqueurs Quantitatifs des Images (BQI) médicales ont le potentiel de compléter, voire de remplacer et surpasser, les biomarqueurs obtenus avec des biopsies dans certains contextes cliniques. Bien que les BQIs n’aient ni la résolution spatiale des images histopathologiques, ni la spécificité moléculaire des analyses génomiques, elles peuvent quantifier la densité et l’hétérogénéité de l’entier des lésions, et ce, de manière non-invasive.
Lay summary

Deux groupes principaux de BQIs existent : les paramètres d’intensité des voxels et les paramètres de texture. Grace à leur facilité d’implémentation et d’interprétation, les paramètres d’intensité ont été largement utilisés pour la prédiction de la survie du patient et réponse au traitement. Les paramètres de texture sont complémentaires et peuvent fournir des mesures précises sur l’architecture locale des tissus lésionnels associée avec la prolifération de cellules cancéreuses, l’angiogenèse ou la nécrose par exemple. Cependant, il existe plusieurs obstacles à leur intégration sereine dans une utilisation clinique.

Notre principal objectif et de proposer des solutions pour améliorer la spécificité, l’interprétabilite et la reproductibilité des BQIs basés sur la texture. Ce projet est innovant non seulement dans le développement de nouveaux BQIs, mais également dans la méthode de validation de ces derniers utilisant une granularité sub-lésionelle. Ces validations seront effectuées à partir de grandes cohortes assemblées au Centre Hospitalier Universitaire Vaudois (CHUV), au travers des collaborations internationales et données publiques. Ces cohortes contiendront des informations locales sur les types de tissus observés grâce à l’utilisation de biopsies stéreotactiques et de traceurs moléculaires. Bien que ces informations soient coûteuses et difficiles à obtenir en routine clinique, elles pourront être utilisées de manière transitoire afin de consolider les liens entre les BQIs et la composition des tissus lésionnels.

Direct link to Lay Summary Last update: 16.08.2018

Responsible applicant and co-applicants


Project partner

Associated projects

Number Title Start Funding scheme
154891 Highly Adaptive Computational Models of Biomedical Tissue in Radiological Images: Digital Tissue Atlases and Correlation with Genomics (MAGE) 01.03.2015 Ambizione


The promises of personalized medicine in oncology have been hindered by its dependence on slow, costly and invasive molecular analyses. In addition, the analyzed tissue samples are most often acquired from either a highly localized tumor region or from the grinding of whole tumor mass, which does not allow to accurately capture molecular heterogeneity [69]. Recent studies in the emerging field of radiomics1 showed strong evidence that Quantitative Imaging Biomarkers (QIB) have the potential to surrogate and even surpass biopsy-based molecular assays in well-defined applicative contexts [2,129]. Whereas such imaging biomarkers neither have the spatial resolution of histopathology nor the molecular specificity of assays, they have the potential to capture the global intralesional heterogeneity of tumor phenotypes in a non-invasive way [114,144].Two main groups of QIBs were proposed in the literature of radiomics: intensity and texture. Thanks to their interpretability and ease of implementation, intensity-based measures were widely used in many applica- tions to predict Overall patient Survival (OS), Disease-Free Survival (DFS) and response to treatment [12,17]. Texture-based imaging biomarkers measure complex dependencies between voxel values at multiple image scales and directions [37]. They can provide precise measurements of local tissue architectures associated with cell proliferation, angiogenesis and necrosis [5, 21, 43]. However, there are many obstacles preventing their serene integration into clinical routine [135]. First, unlike intensity-based biomarkers, texture attributes are mostly informative when forming large groups. Two major consequences of the latter are (i) the challenge of apprehend- ing the relation between multivariate analyses of texture attributes and targeted disease outcomes [115], and (ii) the need for large-scale studies to validate multivariate analyses while respecting ten-to-one ratios between the number of samples and attributes [16]. Second, commonly used approaches for texture analysis suffer from important limitations in terms of robustness, invariances and specificity [36]. Third, although the promise of radiomics relies on quantifying tissue heterogeneity, most studies are aggregating voxelwise imaging measures over entire tumoral volumes allowing neither to characterize distinct tumor habitats nor to reveal the complex relationships between imaging biomarkers and underlying physiological processes.In this proposal, solutions are detailed to overcome the aforementioned limitations. These solutions are justified by recent theoretic analyses detailed in our recent book on Biomedical Texture Analysis (BTA) [33]. Previous work already showed the importance of the identified limitations, where novel trainable and highly- specific texture operators achieved state-of-the-art performance in a variety of applications [24,34,45,49,53, 57,64]. However, further research efforts are required to (i) develop texture operators fulfilling all optimiality criteria defined in [33] and (ii) investigate the importance of the criteria in the particular case of radiomics oncology studies. Second, the project will innovate in the validation of the developed methods by moving from tumor- to habitat- level validation (i.e., seeking for cancer hallmarks [68,81]) based on high-quality, innovative and large datasets gathered at the Lausanne University Hospital (CHUV), international collaborations and public sources. These will include collections of subjects with sub-tumoral local ground truth based on localized stereotactic core biopsies for histopathological and molecular analyses [86, 87] as well as molecular tracers of hallmarks (e.g., cell proliferation, hypoxia, angiogenesis) [88,112]. While being expensive and difficult to use in clinical routine, the aforementioned biomarkers can be used transiently to consolidate the link between QIBs and tumor habitats. The proposal will primarily focus on Non-Small Cell Lung Cancer (NSCLC) [138] and Head and Neck Cancers (HNC) [119] as they are two important cancer types with most data and knowledge available [2,105,106]. They are also best imaged using Computed Tomography (CT) and Positron Emission Tomography (PET) images, where voxel values have a direct and absolute physical interpretation.