select search filters
briefings
roundups & rapid reactions
before the headlines
Fiona fox's blog
Two papers published in Nature Medicine looked at deep-learning analyses, one of CT scans and one of eye scans.
Prof Noel Sharkey, Emeritus Professor of Artificial Intelligence and Robotics, University of Sheffield, said:
“If you were looking for beneficial applications of deep learning, this has got to one of them. It’s the type of task that these learning techniques are cut out for. These are substantial studies but the next step is to ensure that they are used in the most effective way by medical staff and not just followed blindly. That will require a carefully crafted interface to prevent automation bias.”
Prof Nasir Rajpoot, Fellow at The Alan Turing Institute and Professor in Computer Science at University of Warwick, said:
“The research reported in these two separate studies advances the use of modern deep learning technologies applied to routine 3D images of cranial and retinal scans. This is immensely exciting as it can significantly help with making the diagnostic process faster and more objective while being accurate. Results of the two studies will, however, need to be reproduced in larger international multi-centric trials before wider clinical adoption.”
Dr Larisa Soldatova, Reader in Data Science, Goldsmiths University of London, said:
[In reference to the paper ‘Clinically applicable deep learning for diagnosis and referral in retinal disease’]
“This paper is a typical successful example of the application of deep neural networks to medical image analysis. The main result of the paper is an automated classification system for OCT that performs as well, or possibly a bit better, than expert doctors. Few people within AI doubt that such systems can be built, and that they will be of high clinical value.
“Deep Mind are probably the most famous AI company in the world, thanks to their development of AlphaGo which now beats the best humans at Go. Deep Mind have been actively seeking to deploy their machine learning technology (deep learning, reinforcement learning) to medical problems for the NHS, mostly to focus on image analysis. However, as Deep Mind is owned by Google a company with a poor reputation for customer confidentiality, there are understandable deep concerns, see e.g. Powles & Hodson (2017) and the BBC: (www.bbc.co.uk/news/technology-39301901).”
Prof Ross King, Professor of Machine Intelligence, University of Manchester, said:
[In reference to the paper ‘Clinically applicable deep learning for diagnosis and referral in retinal disease’]
“This research shows the power and promise of deep neural networks to transform medical image analyses, however there are a number of technical criticisms that can be made of this paper.
“One is that it is written solely from the point of view of the deep neural network ‘tribe’ of machine learning. We would have appreciated a more balanced approach. For example, consider the following statements: ‘after training on only 14,884 scans’, ‘First, AI (typically trained on hundreds of thousands of examples from one canonical dataset)’. In most machine learning applications 14,884 examples would be considered a lot, and the need to train on hundreds of thousands of examples is typical only of deep neural networks. A good case could be made that if you have hundreds of thousands of examples then deep neural networks is the approach of choice. However, in most real-world problems, such as OCT classification, hundreds of thousands of examples are not available, and it far from clear that deep neural networks is the best approach. Therefore, we would have like to have seen some meaningful comparison with other machine learning approaches. Incorporating such approaches would also make the system more robust. It should also be noted that it is unlikely that human doctors would need to see so many examples to learn how to classify OCTs accurately. We believe that the future is in human experts and AI systems working as a team, enhancing each other cognitive abilities.
“In standard statistics, classification systems produce estimates of class probabilities, then decision theory is used to decide on classes. One criticism of neural networks is that they don’t necessarily produce good estimates of class probabilities. Therefore the statement: ‘The output of our framework can be optimized to penalize different diagnostic errors, and thus for other clinically important metrics’ is somewhat misleading.
“Another standard criticism of neural networks is that they don’t produce explainable results. This is an especially strong criticism in medical applications, where important health related decisions are made. This makes the following statement misleading: ‘Here we created a framework with a structure that closely matches the clinical decision-making process, separating judgements (sic) about the scan itself from the subsequent referral decision.’ Human doctors can, if necessary, output explanations for their classifications. It is notoriously hard to interpret the decision making by neural networks.
“The following statement is technically incorrect: ‘where the true tissue type cannot be deduced from the image, and thus multiple equally plausible interpretations exist.’ The tissue type is abduced, not deduced.”
Prof Duc Pham, Chance Professor of Engineering, University of Birmingham, said:
“These two excellent applications of deep learning do indeed have the potential to improve healthcare by assisting clinicians in making rapid and accurate diagnoses. The emphasis on ‘assisting’ is because the deep learning tools developed will not replace clinicians. Like other forms of machine learning, deep learning is inductive, i.e. it forms general rules and principles from specific training examples. Inductive systems cannot be guaranteed to produce 100% correct results, no matter how many training examples they used or how much training they received. Thus, critical judgments or decisions must always be left to qualified human experts to make.”
Prof Derek Hill, President, Professor of Medical Imaging Science, UCL, said:
“These papers demonstrate the impressive performance that artificial intelligence can now achieve in medical imaging applications. Artificial intelligence methods involving neural networks have been applied to medical imaging applications since the late 1980s but it is only recently with faster computing and more data that they have the potential to impact patient management.
“These papers show how an automatic algorithm can perform comparably to an expert doctor in reporting a 3D image and determining the right referral of a patient to subsequent investigation by a specialist. Despite what the press release suggests, these applications are not automatically providing a diagnosis, but are streamlining a clinical workflow, and assisting with diagnosis. This is still important, but means doctors are still very much involved. A clever aspect of the opthalmology application is the way it handles differences in imaging system separately from differences in pathology.
“For these technologies to impact patients, they will need to be approved or cleared as medical devices. Fortunately regulators internationally have been updating their processes to enable such ‘Software as a Medical Devices’ to be made available in hospitals. This includes the use of AI in medical devices which even 5 years ago would behave been almost unthinkable. Key to the regulatory processes in both the US and Europe is a risk analysis. And in particular whether the device is for treating critical, serious or non-serious conditions, and whether it informs patient management, drives patient management or automatically provides a diagnosis or treatment. The papers in this issue of Nature Medicine would likely initially be made available to inform management, i.e. support the doctor rather than replace them. But as the algorithms become more established and their performance is better characterised, they could be upgraded to automate some medical pathways.
“These papers provide further evidence that artificial intelligence will soon be routinely supporting doctors in streamlining diagnosis and treatment of numerous illnesses.”
Prof Martyn Thomas, Professor of IT, Gresham College, London and Director and Principal Consultant of Martyn Thomas Associates Limited, said:
“I have no expertise in the clinical side of this work, but the paper says that they didn’t focus on diagnosis but on triage, because the problem of prioritising images for review by human experts was more tractable than actual clinical diagnosis. In the Titano paper on neurological events they also say ‘our results on actual diagnostic accuracy were poorer than those of human performance’. In my opinion, the first sentence of the press release greatly exaggerates the results reported by the researchers.
“The work still needs significant validation, as the researchers themselves say: ‘Most importantly, though, will be the active involvement of the medical community to ensure that developed algorithms are intelligently integrated into existing medical practice to solve actual medical problems and that techniques are rigorously validated for clinical efficacy.’
“Nevertheless, in my opinion, the work looks promising and its use in triage might already provide significant benefits to patient outcomes, by rapidly identifying the cases that look most serious so that they get the most urgent expert human assessment.”
* ‘Automated deep-neural-network surveillance of cranial images for acute neurologic events’ by Titano et al. and ‘Clinically applicable deep learning for diagnosis and referral in retinal disease’ by De Fauw et al. were published in Nature Medicine on Monday 13th August.
Declared interests
Prof Noel Sharkey: None received
Prof Nasir Rajpoot: “No conflict of interest.”
Dr Larisa Soldatova: “I do not have any competing interests.”
Prof Ross King: “I do not have any competing interests.”
Prof Duc Pham: “I have no conflict of interest to declare.”
Prof Derek Hill: “I have no conflict of interest.”
Prof Martyn Thomas: “I have no conflicts to disclose.”