Four challenges to overcome for AI to thrive in medical imaging
With increased digitisation in medical clinics, comes increased reliance on AI-enabled tools to more efficiently assess and select diagnostic outcomes using artificial intelligence (AI) algorithms based on machine learning techniques. Here, Dr Christian Stoeckigt discusses what he believes needs to happen in parallel with technological transformation to raise confidence and maintain standards.
Over the past several years we have started to see breakthroughs accelerate in the application of artificial intelligence (AI) in medical imaging. For instance, the digitisation and analysis of mammograms, the use of AI-based algorithms to evaluate digitised prostate slides for low-grade and high-grade tumours and the piloting of digital pathology for cervical screening in cytology labs.
Microsoft’s Project InnerEye has developed machine learning techniques, demonstrating how AI can augment and accelerate clinicians’ ability to perform radiotherapy planning 13 times faster. Meanwhile Google Health is experiencing great strides in AI-enabled tools to predict sight-threatening conditions and improve lung cancer detection, demonstrating how AI can enable transformative diagnostics.
While AI-guided diagnostics in cancer screening programmes are already demonstrating huge progress, the field needs to overcome further regulatory, ethical, and technical challenges to realise its full potential. Here are four key challenges that need to be overcome for the use of AI in medical imaging to thrive:
Education about AI and its positive impact on care is vital for patients to accept its involvement in their diagnosis. In my view, it is the responsibility of healthcare providers to communicate this in collaboration with clinicians. For example, educating patients around how AI-guided diagnostics has the potential to deliver greater accuracy of results and the potential to speed up diagnosis, and ultimately treatment. Speaking to one of our customers recently, they explained that piloting our Genius Digital Diagnostics system had not only helped them to triage patients more effectively, but also to free up capacity in their system to dedicate more time to training cytoscreeners. These types of outcomes, presented simply, are crucial for patients to understand in order to recognise AI’s part in medical diagnosis.
Working in collaboration with radiologists, pathologists and cytotechnologists will be key to improving confidence in AI’s medical image analysis capabilities. Clinicians play a critical role in sourcing unusual cases so that the algorithm constantly evolves. In return, clinicians become more confident in the results that AI provides them. For instance, one of our German customer’s labs, which is piloting our digital cytology system, took a known microcarcinoma case and tested the AI-guided diagnostics with it. In minutes, the AI algorithm had identified a small number of abnormal cells. This gave the clinicians confidence to work side by side with the system.
The knock-on impact is that when AI and digitisation is adopted in medical imaging, it will help clinicians to prioritise their workflow, seek second opinions in near-real time with their peers, and turn around diagnoses much faster for patients.
When I speak to experts in digital pathology, doctors and AI specialists, we all agree we need greater clarity and consensus from governments worldwide to decide on a regulatory approach to AI. There are already some CE marked algorithms for use in medical imaging (such as breast and prostate cancers) that assist with prioritisation and risk stratification. This is a positive step, but we still need further consensus. One of the biggest recent breakthroughs has been the United States FDA’s change in regulations and how it approaches AI, giving more guidance on how AI systems can be trained. This is welcome change that many of us working in the AI field would like to see being replicated across other regions.
A safeguarding framework
It is a reality that humans make mistakes in diagnosis, but we also need to consider what happens when AI is involved in misdiagnosis and navigate questions of accountability. We require a multidisciplinary taskforce to work through these types of questions to build a more holistic approach to safeguard AI systems. Asking cytologists, radiologists, oncologists, experts in ethics and IT and clinicians to come together on this issue would help develop broader perspectives to build these types of protocols.
Where next for AI in medical imaging?
When these four areas are tackled in-depth, we will see wider adoption of AI-driven diagnostics, leading to even higher standards of diagnosis and care for patients and earlier detection of cancers. AI has the power to deliver more effective triaging of patients, so they get the care they need, faster, and providing more personalised medicine for better patient outcomes. Even more exciting is the field of prognostics, where in future AI will be able to predict which people are more likely to develop certain cancers. It is these types of innovations that spur on our research, knowing we have the potential to save more lives from cancer.
 Oktay O, Schwaighofer A, Carter D, Bristow M, Alvarez-Valle J, Nori A, Microsoft Research blog [Internet]. Cambridge: 2020 Nov 30 [cited 2021 Jun 17]. Available from: https://www.microsoft.com/en-us/research/blog/project-innereye-evaluation-shows-how-ai-can-augment-and-accelerate-clinicians-ability-to-perform-radiotherapy-planning-13-times-faster. Accessed 30 June 2021
 Google Health. AI-enabled diagnostics previously thought impossible [Internet]. Mountain View: Google [cited 2021 Jun 17] Available from: https://health.google/health-research/imaging-and-diagnostics/. Accessed 30 June 2021
Author: Dr. Christian Stoeckigt, Head of Scientific Affairs & Medical Education at Hologic