EuSoMII Academy 2017

EuSoMII Academy 2017 High profile challenges in machine learning and artificial intel - Currently, the two key keywords in medicine are: value-based ligence such as the Jeopardy! Match in 2011 and Google Deep health care and artificial intelligence. By implementing these Mind’s triumph at “Go”, have resulted in unprecedented spec- concepts there is an important opportunity to change the way ulation about the end of radiology. Articles in the New Eng- radiology is practiced and the future of radiology can be en- land Journal and JACR by Ezekiel Emanuel in recent months sured. have proclaimed that, “in a few years there may be no special- To operationalize the previously noted concepts, radiology ty called radiology”. Stanford’s Andrew Ng suggested in The needs to focus on 3 main topics, namely: automation, quantifi - Economist that a “radiologist may now be in greater danger of cation / datafication and integration. being replaced by a machine than his own executive assistant”. Firstly automation, because the computer can do a lot of tasks This comes after a prominent West Coast start-up proclaimed using artificial intelligence we are currently performing as a an end to the “wasted protoplasm, which is the radiologist at radiologist. This will allow the radiologist to spend more time the workstation”! Despite these concerns there are numerous on value-adding activities. Secondly quantification / datafica - reasons why radiologists need not fear AI. tion, as there is tons of information in an image that can and, 1. A wide variety of problems in statistics and specifically in in fact, needs to be quantified. Furthermore, free-text reports medicine can be solved best with the creation of a “machine” should be abandoned and be replaced by structured reports. which can provide a simulation or model to discern patterns in The obtained parameters guide the clinician in choosing the a dataset and make predictions. most value-adding treatment for the patient. Also, these param- 2. Machine learning is an advanced statistical technique that eters can be used to develop artificial intelligence algorithms. accomplishes this Finally integration, as the radiologists’ task is to interpret the 3. Interpretation of medical images is much more difficult than important data, advise on the added value of additional diag- Deep Learning/Machine learning experts have anticipated for nostic tests, such as the need for more imaging, pathology, or a variety of reasons and radiologists will not be replaced for a laboratory test, and integrate information to guide clinicians quite a long time in their decisions. 4. There are many problems involving quality, efficiency, and Therefore, automation, quantification / datafication and inte - safety in medicine and medical imaging that can be addressed gration by means of artificial intelligence will ensure the future with machine learning/deep learning approaches quite effec- of radiology in the era of value-based imaging. tively. The application of this approach will have a profound impact on the practice of medicine in the future. Insights Imaging (2018) 9 (Suppl 3):S705 ̶ S708 S707 A3 ̶ Unlocking patterns in medical images with AI for developing imaging biobanks. It focused on recent efforts towards linking imaging biobanks and classical (specimen) Author(s): Ben Glocker biobanks, in the context of a collaboration between the Eu- Disclosure statement ropean Society of Radiology and the BBMRI-ERIC research The author is scientific advisor for Kheiron Medical Technologies infrastructure, implemented as a two-step process. The first Topic: Artificial Intelligence step consists in offering the possibility to describe image col- Abstract category: Scientific lections as part of the BBMRI-ERIC Directory, thus requiring to extend the data model of this database to include descriptive Abstract items concerning medical image collections. A second step is Artificial intelligence has the potential to address a major crisis also envisaged, involving the design of information models for in medical imaging. With an ever increasing complexity, vol- describing individual image datasets, as well as important de- ume of data and economic pressure the interpretation of med- rived data, such as imaging biomarkers. The design of these in- ical images pushes human abilities to the limit. According to formation models is envisaged as an extension of the MIABIS The Royal College of Radiologists, in the UK alone there are / OMIABIS model (Minimum Information About BIObank more than 300,000 patients waiting for more than a month for data Sharing). Emphasis was specifically put on the additional their imaging results. For some areas of image-based diagnosis possibilities resulting from the use of semantic web technolo- error rates of up to 30% have been reported. gies (i.e. ontology-based approach), seen as the most appropri- The aim of our research at the BioMedIA group at Imperial ate to ensure linking and managing consistency with data other College London is to develop machines capable of analysing domains (linked data, federated semantic repositories). and interpreting medical scans with super-human performance. We utilise cutting-edge deep learning technology to automati- cally extract clinically useful information from images which A5 ̶ Deep learning: basic principal otherwise would be difficult or impossible to retrieve. We have Author(s): Bram van Ginneken successfully applied this technology to analyse brain and car- Disclosure statement diac images, where our computational tools enable automatic, Topic: Image analysis quantitative assessment of anatomical structures and pathol- Abstract categories: Educational ogies such as brain lesions. Our lab has recently presented a novel approach to segment brain tumors with unprecedented Abstract accuracy which was ranked top at the international Multimodal For more than half a century, researchers have been program- Brain Tumor Segmentation Challenge (BraTS) 2017. ming computers to detect abnormalities en make predictions Despite the recent successes and promising results, we still from medical images. Until a few years ago, the standard ap- face major challenges in deploying this technology in clinical proach was to think carefully about what characteristic num- practice. Future research needs to focus on making machine bers (features) could be computed from the image. A set of learning methods more robust and trustworthy. We observe such numbers was then mapped to a prediction using a training significant degradation of performance on new data and auto - data set of images with known lesions or outcome, and a classi- matic predictions come with limited interpretability. We pres- fier, or statistical model. This field is known as machine learn - ent some recent ideas and preliminary results to tackle these ing. Deep learning is fundamentally different because now the challenges. images are directly used as input to a neural network. The step is designing the right features to be computed is omitted. The A4 ̶ Image sharing and biobanks network is trained using a large data set of images with known lesions or outcome. Training here means endlessly fiddling Author(s): Bernard Gibaud with the internal weights of the network until the network pre- Disclosure statement dicts the correct outcome for new images. Topic: Imaging biobanks, semantic web Since 2012, deep learning has been applied to over 300 tasks Abstract categories: Scientific in medical image analysis. I highlighted some recent results in Abstract the analysis of fundus photographs for the detection of diabetic Data sharing has become a primary issue in many domains retinopathy, in computational pathology for the detection of of modern life, as a result of the progress of communication metastases from lymph node biopsies and in chest CT analysis and storage technology and spurred by the huge possibilities for the detection of nodules and the prediction of the presence offered by recent data mining technology (big data mining, of lung cancer. In all these applications, deep networks per- deep learning). Given the prominent role of medical imaging form comparably to human experts. in modern medicine and especially in precision medicine, the The results I presented demonstrate that with today’s technolo- development of imaging biobanks appears as critically impor- gy it is possible to automate visual tasks that humans can per- tant for biomedical research towards achievement of precision form after an extended period of training. However, there are medicine. still few such automated systems validated and widely used in The presentation first summerized the most basic motivations practice. Moreover, clinicians perform thousands of medical S708 Insights Imaging (2018) 9 (Suppl 3):S705 ̶ S708 image analysis tasks in their daily work. Thus, there is much Abstract work to do in developing these systems, and this endeavor has Recent years has shown an explosive growth in the use of Ar- a high potential for making health care more affordable and tificial Intelligence (AI) and Deep Learning not in the least for more effective. medical applications. These new technological developments have started a whole new discussion on privacy of data pro- A6 ̶ Integrating data in the Structured Report cessed by these computerized systems. Especially those appli- cations in health care demand a high level of patient privacy Author(s): Daniel Pinto dos Santos, and patient data security. The actual ownership of medical data Disclosure: None is also part of this discussion where we can have different sit- Topic: Structured reporting, interoperability, computer applications uations with original, de-identified, anonymized and processed Abstract categories: Educational data. Questions like what data is still personal data for an indi- vidual patient or participant in a clinical trial and who actual- Abstract ly owns the data that is produced with self-learning computer Although structured reporting is still not widespread in clinical systems. This educational course will discuss these issues with routine, it has been a major topic in the radiological commu- respect to ownership, intellectual property and the different as- nity for quite a while. With the development and publication pects that are involved in this when moving towards the era of the IHE MRRT profile, actors and transactions in structured of AI. reporting have been very well described. The profile allows for data from different sources to be incorporated in the structured report (e.g. via HL7, from RIS, PACS or EMR). Moreover, A9 ̶ Transitioning to AI-Powered Highly Automated Ra- template modules can be integrated in other templates and diology structured reporting templates themselves can already have im- Author(s): J. Raymond Geis plicit data integrated such as knowledge about categorization Disclosure statement: No conflict of interest, and nothing to of certain pathologies. In this talk an overview of the existing disclose possibilities to enrich structured reports with data from various Topic: Radiology AI; Imaging Informatics; Radiology practice sources will be given, as well as an outlook where data coming Abstract categories: Imaging Informatics, Radiology practice from structured reports can be used for other applications. Abstract A7 ̶ Role of image computing in radiology - opportunity In the automobile industry, self-driving cars are known as or threat? Highly Automated Vehicles (HAVs). Work on HAVs has been going on for decades. Today, the general world still drives cars Author(s): Paul Suetens much like we did years ago, though now our cars will auto- Disclosure: None matically brake for objects in front of us, or will warn us if we Topic: AI, Machine learning, Image computing stray from our lane. These are narrow AI tools, each designed Abstract categories: Educational to perform a focused task. Abstract Advances in computing power, algorithms and digital data The 1970s was the decade that “computed imaging” radically provide the basis for a future of highly-automated radiology changed the field of radiology. Today “image computing” has (HAR). Similar to cars, we have just entered a phase that will become sufficiently mature to have a similar influence on this probably last for decades, where we will still “drive” radiol- discipline. In this talk the principles of image computing will ogy similar to how we do today, but with the addition of our be shortly summarized and its potential will be illustrated on own narrow AI tools. These narrow AI tools broadly fit three clinical examples of computer-assisted detection, screening, groups: quantitative measurements, evidence-based diagnosis, early · Radiomics, screening for disease, computer-assisted detection outcome prediction of therapy, and 3D visualization. Obvi- · Knowledge management and Natural Language Processing ously, exploiting this new technology is a logical evolution in (NLP) the context of value-based health care. It should therefore not · Protocol and workflow management – aka “Protocolomics” be neglected, but instead, be considered as an opportunity and It is surprisingly easy to write a computer program to do a radi- adopted by radiologists. ology-AI task on well-defined, highly curated, local data. It is difficult, however, to build an AI product that will work on di - A8 ̶ Who Controls Patient Data in the Era of A.I.: man or verse, heterogeneous input data. On top of this are ethical and machine? bias issues, how to handle rare but “do-not-miss” cases, and how to verify that a “learning algorithm” is learning the correct Author(s): Peter M.A. van Ooijen, MSc, PhD, CPHIT things, and becoming better rather than worse for all situations. Disclosure: None Topic: Artificial Intelligence, privacy, data security Abstract categories: Educational EuSoMII Annual Meeting 2018 rd 3 November 2018-EMC, Rotterdam For further info please contact Mrs. Paola Rinaldi prinaldi.job@gmail.com http://www.deepdyve.com/assets/images/DeepDyve-Logo-lg.png Insights into Imaging Springer Journals

EuSoMII Academy 2017

Insights into Imaging , Volume 9 (3) – May 23, 2018
Free
4 pages