Data infrastructures for AI in medical imaging: a report on the experiences of five EU projects.

Artificial intelligence (AI) is transforming the field of medical imaging and has the potential to bring medicine from the era of 'sick-care' to the era of healthcare and prevention. The development of AI requires access to large, complete, and harmonized real-world datasets, representative of the population, and disease diversity. However, to date, efforts are fragmented, based on single-institution, size-limited, and annotation-limited datasets. Available public datasets (e.g., The Cancer Imaging Archive, TCIA, USA) are limited in scope, making model generalizability really difficult. In this direction, five European Union projects are currently working on the development of big data infrastructures that will enable European, ethically and General Data Protection Regulation-compliant, quality-controlled, cancer-related, medical imaging platforms, in which both large-scale data and AI algorithms will coexist. The vision is to create sustainable AI cloud-based platforms for the development, implementation, verification, and validation of trustable, usable, and reliable AI models for addressing specific unmet needs regarding cancer care provision. In this paper, we present an overview of the development efforts highlighting challenges and approaches selected providing valuable feedback to future attempts in the area.Key points• Artificial intelligence models for health imaging require access to large amounts of harmonized imaging data and metadata.• Main infrastructures adopted either collect centrally anonymized data or enable access to pseudonymized distributed data.• Developing a common data model for storing all relevant information is a challenge.• Trust of data providers in data sharing initiatives is essential.• An online European Union meta-tool-repository is a necessity minimizing effort duplication for the various projects in the area.

Position of the AI for Health Imaging (AI4HI) network on metadata models for imaging biobanks.

A huge amount of imaging data is becoming available worldwide and an incredible range of possible improvements can be provided by artificial intelligence algorithms in clinical care for diagnosis and decision support. In this context, it has become essential to properly manage and handle these medical images and to define which metadata have to be considered, in order for the images to provide their full potential. Metadata are additional data associated with the images, which provide a complete description of the image acquisition, curation, analysis, and of the relevant clinical variables associated with the images. Currently, several data models are available to describe one or more subcategories of metadata, but a unique, common, and standard data model capable of fully representing the heterogeneity of medical metadata has not been yet developed. This paper reports the state of the art on metadata models for medical imaging, the current limitations and further developments, and describes the strategy adopted by the Horizon 2020 "AI for Health Imaging" projects, which are all dedicated to the creation of imaging biobanks.

A quality control portal for sequencing data deposited at the European genome-phenome archive.

Since its launch in 2008, the European Genome-Phenome Archive (EGA) has been leading the archiving and distribution of human identifiable genomic data. In this regard, one of the community concerns is the potential usability of the stored data, as of now, data submitters are not mandated to perform any quality control (QC) before uploading their data and associated metadata information. Here, we present a new File QC Portal developed at EGA, along with QC reports performed and created for 1 694 442 files [Fastq, sequence alignment map (SAM)/binary alignment map (BAM)/CRAM and variant call format (VCF)] submitted at EGA. QC reports allow anonymous EGA users to view summary-level information regarding the files within a specific dataset, such as quality of reads, alignment quality, number and type of variants and other features. Researchers benefit from being able to assess the quality of data prior to the data access decision and thereby, increasing the reusability of data (https://ega-archive.org/blog/data-upcycling-powered-by-ega/).

Beacon v2 and Beacon networks: A "lingua franca" for federated data discovery in biomedical genomics, and beyond.

Beacon is a basic data discovery protocol issued by the Global Alliance for Genomics and Health (GA4GH). The main goal addressed by version 1 of the Beacon protocol was to test the feasibility of broadly sharing human genomic data, through providing simple "yes" or "no" responses to queries about the presence of a given variant in datasets hosted by Beacon providers. The popularity of this concept has fostered the design of a version 2, that better serves real-world requirements and addresses the needs of clinical genomics research and healthcare, as assessed by several contributing projects and organizations. Particularly, rare disease genetics and cancer research will benefit from new case level and genomic variant level requests and the enabling of richer phenotype and clinical queries as well as support for fuzzy searches. Beacon is designed as a "lingua franca" to bridge data collections hosted in software solutions with different and rich interfaces. Beacon version 2 works alongside popular standards like Phenopackets, OMOP, or FHIR, allowing implementing consortia to return matches in beacon responses and provide a handover to their preferred data exchange format. The protocol is being explored by other research domains and is being tested in several international projects.

The European Genome-phenome Archive in 2021.

The European Genome-phenome Archive (EGA - https://ega-archive.org/) is a resource for long term secure archiving of all types of potentially identifiable genetic, phenotypic, and clinical data resulting from biomedical research projects. Its mission is to foster hosted data reuse, enable reproducibility, and accelerate biomedical and translational research in line with the FAIR principles. Launched in 2008, the EGA has grown quickly, currently archiving over 4,500 studies from nearly one thousand institutions. The EGA operates a distributed data access model in which requests are made to the data controller, not to the EGA, therefore, the submitter keeps control on who has access to the data and under which conditions. Given the size and value of data hosted, the EGA is constantly improving its value chain, that is, how the EGA can contribute to enhancing the value of human health data by facilitating its submission, discovery, access, and distribution, as well as leading the design and implementation of standards and methods necessary to deliver the value chain. The EGA has become a key GA4GH Driver Project, leading multiple development efforts and implementing new standards and tools, and has been appointed as an ELIXIR Core Data Resource.

Verbing nouns and nouning verbs: Using a balanced design provides ERP evidence against "syntax-first" approaches to sentence processing.

In this event-related potential (ERP) study we reevaluate syntax-first approaches to sentence processing by implementing a novel paradigm in French that includes correct sentences, pure syntactic category violations, lexical-semantic anomalies, and combined anomalies. Our balanced design systematically controlled for target word (noun vs. verb) and the context immediately preceding it. Group results from 36 native speakers of Quebec French revealed that, up to 300 ms, ERPs elicited by syntactic category violations were comparable with ERP responses to correct sentences, showing that there is no early activation reflecting syntactic category identification. Instead, in response to all anomalous conditions, we observed an N400 followed by a P600. Combined anomalies yielded additive effects of syntactic category and lexical-semantic anomalies on the N400, and a large P600 effect similar to the one observed in the pure syntactic condition. These results provide strong evidence against the hypothesis that (i) syntactic categories are processed first, and (ii) that syntactic category errors "block" lexical-semantic processing. Further, the N400 effect in response to pure syntactic category violations reflects a mismatch detection between a predicted word-stem and the actual target. This mechanism takes place simultaneously (and potentially in parallel) with lexical-semantic processing. In contrast, an interaction of syntax and semantics for the P600 reveals that the same neurocognitive resources are recruited for syntactic and semantic integration, both promoted by the implementation of an acceptability judgement task in our design. Additional analyses of individual data complemented these observations: during sentence processing, participants did not rely on one single cognitive mechanism reflected by either the N400 or the P600 effect but on both, suggesting that the biphasic N400-P600 ERP wave can indeed be considered to be an index of phrase-structure violations in most individuals, at least if they are realized on content words.

Growing Random Forests reveals that exposure and proficiency best account for individual variability in L2 (and L1) brain potentials for syntax and semantics.

Late second language (L2) learners report difficulties in specific linguistic areas such as syntactic processing, presumably because brain plasticity declines with age (following the critical period hypothesis). While there is also evidence that L2 learners can achieve native-like online-processing with sufficient proficiency (following the convergence hypothesis), considering multiple mediating factors and their impact on language processing has proven challenging. We recorded EEG while native (n = 36) and L2-speakers of French (n = 40) read sentences that were either well-formed or contained a syntactic-category error. a lexical-semantic anomaly, or both. Consistent with the critical period hypothesis, group differences revealed that while native speakers elicited a biphasic N400-P600 in response to ungrammatical sentences, L2 learners as a group only elicited an N400. However, individual data modeling using a Random Forests approach revealed that language exposure and proficiency are the most reliable predictors in explaining ERP responses, with N400 and P600 effects becoming larger as exposure to French as well as proficiency increased, as predicted by the convergence hypothesis.

The priming of priming: Evidence that the N400 reflects context-dependent post-retrieval word integration in working memory.

Which cognitive processes are reflected by the N400 in ERPs is still controversial. Various recent articles (Lau et al., 2008; Brouwer et al., 2012) have revived the idea that only lexical pre-activation processes (such as automatic spreading activation, ASA) are strongly supported, while post-lexical integrative processes are not. Challenging this view, the present ERP study replicates a behavioral study by McKoon and Ratcliff (1995) who demonstrated that a prime-target pair such as finger - hand shows stronger priming when a majority of other pairs in the list share the analogous semantic relationship (here: part-whole), even at short stimulus onset asynchronies (250ms). We created lists with four different types of semantic relationship (synonyms, part-whole, category-member, and opposites) and compared priming for pairs in a consistent list with those in an inconsistent list as well as unrelated items. Highly significant N400 reductions were found for both relatedness priming (unrelated vs. inconsistent) and relational priming (inconsistent vs. consistent). These data are taken as strong evidence that N400 priming effects are not exclusively carried by ASA-like mechanisms during lexical retrieval but also include post-lexical integration in working memory. We link the present findings to a neurocomputational model for relational reasoning (Knowlton et al., 2012) and to recent discussions of context-dependent conceptual activations (Yee and Thompson-Schill, 2016).

Searching High and Low: Prosodic Breaks Disambiguate Relative Clauses.

During natural speech perception, listeners rely on a wide range of cues to support comprehension, from semantic context to prosodic information. There is a general consensus that prosody plays a role in syntactic parsing, but most studies focusing on ambiguous relative clauses (RC) show that prosodic cues, alone, are insufficient to reverse the preferred interpretation of sentence. These findings suggest that universally preferred structures (e.g., Late Closure principle) matter far more than prosodic cues in such cases. This study explores an alternative hypothesis: that the weak effect of prosody might be due to the influence of various syntactic, lexical-semantic, and acoustic confounding factors, and investigate the consequences of prosodic breaks while controlling these variables. We used Spanish RC sentences in three experimental conditions where the presence and position (following the first or second noun phrase) of prosodic breaks was manipulated. The results showed that the placement of a prosodic break determined sentence interpretation by changing the preferred attachment of the RC. Listeners' natural preference for low attachment (in the absence of break) was reinforced when a prosodic break was placed after the first noun. In contrast, a prosodic break placed after the second noun reversed the preferred interpretation of the sentence, toward high attachment. We argue that, in addition to other factors, listeners indeed use prosodic breaks as robust cues to syntactic parsing during speech processing, as these cues may direct listeners toward one interpretation or another.