Artificial intelligence

Medical Imaging generates large amounts of data commonly assessed visually using scientific evidence and expert knowledge combined with clinical patient-specific information. Quantitative methods have also been developed in today’s multi-modality environment, providing morphologic, functional and biological information. Recent advances in data mining of quantitative image features together with powerful image analytic tools have lead to what is known today as “radiomics”. Naturally, radiomics analytic techniques have been welcomed with both enthusiasm and scepticism. The process used in radiomics involves the identification of vast arrays of quantitative parameters within digital images.. The major challenge is to integrate radiomics data with clinical, pathological, and genomic information to decode the different types of tissue biology and specificities of the disease within a patient in a new paradigm that we introduce in this talk as "radiomics+". Imaging is indeed not the only field roaming the uncharted territory of large medical datasets. Pathology, liquid biopsy techniques or genetic analyses all provide an increasing number of molecular biomarkers. Oncology is the large port of entry of this trend given the need to tailor an increasingly diverse array of targeted therapies to the specificities of each patient and the history of each cancer within each individual. Many institutions are in the process of developing analytic personalised oncology programs (APOs). Through this prism and early local and wider APO experience, this talk will endeavour to give credit to both the traditional expert visual assessment and the radiomics approach and analyse the challenges ahead.

Hybrid imaging lends itself to the measurement of a large number of radiomic features. These include so-called handcrafted features and deep features derived either from the native images or from parametric images. This wealth of image-based information has a great potential for decision making using modern AI approaches. Yet, it is likely that if images are necessary for in vivo probing of biological mechanisms, they are not sufficient for accurate prediction of the patient outcome that depends on a number of other parameters. This is why radiomic data have to be enriched with other relevant omic or clinical data, yielding the concept of holomics. The integration of that huge diversity of data into accurate and robust models for guiding precision medicine raises a number of new challenges that will be presented and illustrated. Addressing these challenges may actually need rethinking the way we conduct research in imaging as will be discussed. In addition, the overwhelming trend towards AI-based data analysis should not hide the crucial importance of both the quality and the relevance of the input data, whatever they are. Only a clever and thorough use of sensitive and specific images reflecting biological mechanisms in an understandable way and combined with other data that also express phenomena to be accounted for will allow us to approach the truth. In that respect, AI has to be combined with human intelligence to make the most of imaging and non-imaging data in the context of precision medicine.

The developments of imaging biobanks and cloud-based data storage services have radically changed the way we deal with communication and data management in our daily life. On-line storage of medical images is not new; several vendors have offered such services for decades already as part of their commercial solutions. What has really changed these recent years is the emergence of such services for the wide public offering very attractive solutions at a very low cost. In medical applications, however, such systems must comply with strict regulations and guidelines geared toward protecting patient confidentiality and data security. Medical imaging is becoming a major component of the data required in every medical decision in diagnostic, assessment of treatment response, follow-up of disease recurrence and in support for therapeutic and surgical interventions. The wealth of data acquired in clinical routine these days is overwhelming and has not been apprehended yet. The main limiting factor of the development of these new analysis techniques is the lack of sufficiently large sets of structured and well-documented imaging data. There are also major difficulties in the ability to collect these large sets of imaging data due to restrictive regulatory constraints and data protection rules that prevent the usage and exploitation of medical data without formal patient approval. Our presentation will focus on the specific issue of gathering and collecting medical images for the development of large Big-Data repositories for scientific research and review the current challenges that prevent their wide development today.

In their recent editorial in Academic Radiology, Cohan and Davenport refer to radiologist “burnout” and having reached a “tipping point”. They suggest that despite improvements in PACS and EMR's, “Radiologists are still being told to work faster as the screws continue to tighten; more images, greater case volume, increasing complexity and less time to do the work”. Radiologists are increasingly asked to perform quantitative analysis on complex dynamic studies such as prostate and breast MRI, analyse multi-parametric imaging from MRI, PET, CT, and to follow new guidelines for lung cancer and other screening studies. Deep learning represents a fundamentally different approach to the development of algorithms for image acquisition, quantitative analysis, and interpretation based on learning by example from large image sets. It offers numerous advantages over more “traditional” Computer Aided Design approaches including decreased time, and less specialised medical imaging expertise required for development as well as the potential for continuous and personalised refinement of algorithms. In fact, Deep Learning may actually have its greatest initial success in solving non-image related challenges such as image quality, workflow efficiency, improved communication and patient safety. This technology, however, is also fraught with limitations including the requirement for large amounts of annotated data, regulatory, medicolegal, and relative brittleness with regard to lack of generalizability from a few to a multitude of different scanners. Overall, despite the challenges, Deep Learning will undoubtedly have a major impact in the next several years on positively resetting radiology’s current “tipping point”.

The combination of big data and artificial intelligence are dramatically increasing the possibilities for prevention, cure and care, and are changing the landscape of the healthcare system. Biomedical imaging data will play a central role in this revolution. In this presentation, I will show examples of possible large benefits of big data analytics of imaging, genetic and clinical data in dementia and oncology. Both conventional machine learning techniques, such as radiomics for tumour characterisation, and deep learning techniques that directly learn from the imaging data will be addressed. Also, the concept of deep imaging, full integration of medical imaging and machine learning, will be discussed. Finally, I will address the challenges of how to successfully integrate these technologies into the daily clinical workflow.

An overview of how machine learning technologies may play a role in the workflow and task automation of radiologists, from appropriateness criteria to image acquisition, to image perception tasks and report generation, this talk will look at the entire ecosystem of diagnostic radiology and AI in the coming years.