In the nine years since AlexNet gave birth to the deep learning era, artificial intelligence (AI) has made significant technological progress in medical imaging, with more than 80 deep learning algorithms Officially recognized Since 2012, it has been used by the US FDA for clinical application of image detection and measurement. A survey in 2020 found that more than 82% of image providers believe that artificial intelligence will Improve diagnostic imaging In the next 10 years, the artificial intelligence market in medical imaging is expected to 10 times increase In the same period.
Despite the optimistic outlook, artificial intelligence still cannot be widely used in radiology.A kind 2020 survey A study by the American College of Radiology (ACR) showed that only about one-third of radiologists use AI, mainly to enhance image detection and interpretation; among two-thirds of people who do not use artificial intelligence, most Most people said that they did not see any benefit.In fact, most radiologists will say that artificial intelligence has Is not Change image reading or improve their practices.
Why is there such a huge gap between the theoretical utility of artificial intelligence and its practical application in radiology? Why has artificial intelligence not fulfilled its promise in radiology? Why have we not “arrived” yet?
The reason is not because the company did not try to innovate. This is because they tried to automate the work of radiologists-but failed, burning a large number of investors, making them unwilling to fund other projects aimed at transforming the theoretical utility of artificial intelligence into real-world use cases.
Artificial intelligence companies seem to have misunderstood Charles Friedman’s basic principles theorem Biomedical informatics: It is not that computers can do more things than humans; a person can do more than just a person using a computer. Creating this kind of human-machine symbiosis in radiology will require artificial intelligence companies to understand:
- The clinical proficiency of radiologists and the construction of algorithms provide context for the computer
- Discrete tasks and build tools for workflow to automate rote or tedious tasks
- User experience and build an intuitive interface
These features are provided together as a unified cloud-based solution that will simplify and optimize radiology workflows while enhancing the intelligence of radiologists.
history class
Modern deep learning began in 2012, when AlexNet won the ImageNet challenge, which led to the resurgence of what we think of today as AI. With the full resolution of the image classification problem, artificial intelligence companies decided to apply their algorithm to the image that has the greatest impact on human health: radiographs. These post-AlexNet companies can be considered as divided into three generations.
When the first generation entered the field, they assumed that AI expertise was sufficient to achieve commercial success, so they focused on building early teams with algorithmic knowledge. However, the team greatly underestimated the difficulty of acquiring and labeling medical imaging data sets large enough to train these models. Without sufficient data, these first-generation companies will either fail or have to give up radiology.
The second generation corrects the failure of its predecessors by establishing data partnerships with academic medical centers or large private medical groups. However, these startups have encountered the dual problem of integrating their tools into radiology workflows and building business models around them. Therefore, they finally built functional features without any commercial appeal.
The third generation of artificial intelligence companies in the field of radiology realized that in addition to algorithms and data, success also requires an understanding of the radiology workflow. These companies are largely focused on the same use case: classification. Their tool sorts the images according to the urgency of the patient, thus sorting the radiologist’s workflow without interfering with the execution of the work.
The third-generation radiology workflow solution is a positive advance and shows that there is a path to take, but in addition to classification and reordering of work lists, artificial intelligence can do more. So where should the next wave of artificial intelligence go in the field of radiology?
Go with the flow
So far, artificial intelligence has proven its value in its ability to handle asynchronous tasks such as image classification and detection. What’s more interesting is to enhance the potential of real-time image interpretation by providing computer context, allowing it to work with radiologists.
Many aspects of the radiologist’s workflow need to be improved, and the AI-based context can be optimized and simplified. These include, but are certainly not limited to: setting the radiologist’s preferred image suspension protocol; automatically selecting the appropriate report template for the case; ensuring that the radiologist dictates the correct part of the report; and there is no need to repeat image measurements for the report.
Separately, the shortcut to optimizing any of these workflow steps—micro-optimization—has little impact on the entire workflow. However, the overall summary of these micro-optimizations will have a considerable collective impact on the radiologist’s workflow.
In addition to the impact on radiology workflows, the concept of “micro-optimization outlines” makes feasible and sustainable businesses; and it is difficult, if not impossible, to build a business around tools that optimize only one of the steps .
Radiology Tool of Thinking
In other areas of software development, we are witnessing a renaissance of “thinking tools” (technologies that expand the human mind), and in these areas, creating products that improve decision-making and user experience is a bet. The adoption of this idea in the healthcare sector has been slower because computers and technology have failed to improve usability and workflow, and there is still a lack of integration.
With the emergence of new applications for image screening and diagnosis, the number and complexity of medical images have continued to increase; however, the total number of radiologists has not increased at the same rate. Therefore, the continuous expansion of medical imaging requires better thinking tools. Without them, when we cannot read all the images generated, we will eventually reach a tipping point and patient care will be affected.
The next wave of artificial intelligence must address the workflow of real-time interpretation of radiology, and we must embrace it when the technology arrives. No single function can solve this problem. Only the micro-optimization summary of continuous high-speed delivery through the cloud can solve this problem.
Photo: Metamorworks, Getty Images
