AI technologies for medical imaging have grown extensively in recent years and have grown extensively in recent years, mainly driven by advances in GPUs and machine learning [1]. With regard to the field of medical imaging, the main objective of a radiologist is to identify patterns, as well as correlations between the presence of these patterns and clinical, laboratory, histopathological and evolutionary findings in several chronic degenerative diseases, corroborating for a shorter time until the diagnosis and rapid institution of an early treatment (obviously, this is an overly simplified view of the work of a radiologist) [1,2]. The potential of AI to reduce the workload and the bias related to human capacity (fatigue, mental health, sleep quality and numerous other factors which contribute to the decrease in the radiologist’s diagnostic accuracy), bring a light at the end of the tunnel that lead to widespread speculation that AI will solve many difficulties in the medical imaging industry [1].
However, despite the hype and potential, there was little uptake of AI in integrated health care, the main reason being attributed to lack of access to useful and representative data. According to Gartner, in 2020, 805 of AI projects did not become real production; of those that did, 40% were not profitable. With over $50 billion of global investment in AI in 2020, markets still don’t seem to be embracing AI, and those that have are not seeing a good ROI [3].
So what is the real issue with data access? Turning data into something accessible is not as simple as it sounds. The data, above all, are sensitive information about the health and clinical condition of patients and research participants, spread across various locations (websites and systems), inaccessible, unlabeled, incorrectly stored and biased in certain cases. The use and access of data require several aspects, from the management of ethical, clinical, legal and regulatory aspects whose main objectives are the protection of patients and research participants. One conducted by Kruse et al. show that Hospitals have lower cybersecurity of their data compared to industries [4]. Findings that corroborate the significant increase in cyber attacks by 67% in Europe [5] and 55% in the USA [6] in the period 2017/2018. In this way, the need for mechanisms to prevent the misuse of these data is of fundamental importance, however, they result in a need for standardization regarding the treatment and use of these data by various health institutions and companies. Even the big, well-funded AI developers still struggle to get data. And it’s extremely time consuming.
As an example, a large North American hospital was excited about the potential of an AI platform that would improve patient care at its facility. As they planned to build a prototype, they found that the data needed to build the AI models was spread across 20 different locations. The time with all interested parties for the formalization of all regulatory aspects, eventual review and standardization of databases (label review, de-identification and all processes related to the protection of patients, research participants, centers and professionals involved) [6]. Retrieving that data would be too complex and time-consuming. The project had to be canceled. In this context, there is an imminent need for companies that can create standardized data access in accordance with current cybersecurity standards and in compliance with the ICH/E2A guide for the management and protection of clinical data [7].
At Gradient, we created a business model and developed a platform to address exactly these limitations. We want to enable AI developers and researchers to have instant access to millions of diverse, high-quality medical imaging datasets from hundreds of hospitals, clinics, and research centers around the world, labeled with no need to this data, which has the following implications: a) diverse and reproducible AI models using real-world data. b) reduction in the possibility of developing a flawed AI system in medical imaging with regard to medical diagnosis, contributing to better management and treatment of various chronic degenerative diseases.
References
[1] M. J. Willemink, W. Koszek, H. Hugh, R. M. Summers, D. Rubin, and M. Lungren, “Preparing Medical Imaging Data for Machine Learning | Radiology,” Radiology, Feb. 2020, Accessed: Mar. 01, 2022. [Online]. Available: https://pubs.rsna.org/doi/full/10.1148/radiol.2020192224
[2] Saw SN, Ng KH. Current challenges of implementing artificial intelligence in medical imaging. Phys Med. 2022 Aug;100:12-17. doi: 10.1016/j.ejmp.2022.06.003. Epub 2022 Jun 14.
[3] McKinsey, “Global AI Survey 2021 – Desktop,” Dec. 2021. Accessed: Oct. 27, 2022. [Online]. Available: http://ceros.mckinsey.com/global-ai-survey-2020-a-desktop-3-1
[4] Holloway S. Irish cyberattack provides wake-up call for European imaging IT. Available at: https://www.auntminnieeurope.com/index.aspx?sec=sup&sub=pac&pag=dis&ItemID=620205. Accessed May 25, 2021:
[5] Vaidya A. Report: Healthcare data breaches spiked 55% in 2020. Available at: https://medcitynews.com/2021/02/report-healthcare-data-breaches-spiked-55- in-2020/. Accessed February 17, 2021: MedCityNew.
[6] O. Gonzalez, “Cybercrime in the US jumped by 55% in the past two years,” CNET. https://www.cnet.com/news/privacy/cybercrime-in-the-us-jumped-by-55-in-the-past-two-years/ (accessed Oct. 28, 2022).
[7] ICH (2005a) Draft consensus guideline – revision of the ICH guideline on clinical safety data management: data elements for transmission of individual case safety reports E2B(R3). ICH Secretariat, Geneva.
