Introduction

How do you feel knowing that your X-rays may be interpreted by a computer program? Surprised? Continue reading about how AI is starting to assist medical imaging. Perplexed? Learn more here about the principles of AI-assisted medical imaging. Troubled?  Discover the important safety and ethical considerations of using AI in medical imaging. 

In the absence of AI, radiology relies on human specialists to interpret, annotate, and report medical scans. In many cases, this workflow can be exhausting and prone to errors.¹ For hospitals with ever-increasing demands on imaging, radiologists consistently have to tackle new studies. At the same time, they are tasked with handling requests from clinicians, which can distract their thought process and lead to costly errors. Radiologists, being human, are also subject to various cognitive biases.² For the rising backlog of scans, felt acutely in hospitals throughout the UK, AI has been proposed as a possible solution. 

The advent of Ai in medical imaging

The definition of AI is not unequivocal. For the purpose of this article, while hard coding (explicitly stating the rules of the program) is included in the scope of AI, the main focus will be software able to solve problems by learning from past solutions.³ 

AI has been used in various disciplines since its conception in the 1950s. A decade later, radiology researchers made their first attempts to design computer programs to assess the probability of a disease from each scan.⁴ However, those early efforts did not yield promising outcomes due to limitations in computing power, data availability, and imaging processing tools.³ Following the development of computational tools and advanced algorithms in the 1980s, however, there was a new appetite for AI-assisted medical imaging.

Milestones in Ai-assisted medical imaging

Prerequisites fulfilled

For an AI system to handle a given task effectively, there has to be a computer powerful enough to carry out the model training process, sufficient data to train the model, and an algorithm suitable for the situation. The first was achieved by rapid enhancements in the computational power of graphics processing units (GPU) in new computer models during the second half of the twentieth century. Radiology data has also accumulated since electronic health records (EHR) were gradually adopted by radiology departments between the 1970s to 1990s.⁵ The last requirement was fulfilled in the 1980s when various advanced Machine Learning (ML) algorithms were developed.⁶

Machine Learning (ML) is a subset of AI that primarily leverages existing data to “learn” patterns, allowing models to predict or determine outcomes in new situations. There are many ML algorithms, one of the most important being the artificial neural network (ANN). It learns by adjusting the connections between nodes of different layers, simulating the learning process between neurones in the human brain. This architecture makes ANN appropriate for image processing because it can identify various features, such as extracting boundaries from the images and thus extracting shapes from boundaries.

The emergence and growth of computer-aided diagnosis (CAD)

With all the prerequisites in data, hardware and algorithms fulfilled, computer-aided diagnosis (CAD) was born. In comparison to the early attempts to use AI in radiology, CAD aims to aid radiologists, rather than substitute them. CAD as a model is trained using radiology databases (including a patient’s medical history, diagnosis, and scans). The trained model can process new scans and other important input data to suggest likely abnormalities. These suggestions can assist radiologists by identifying subtle structures or providing second opinions.

The first CAD research was carried out in the mid-1980s. Early investigations mostly focused on cardiovascular problems and cancer.⁷ Similar to drugs and medical devices, CAD needs clinical trials and regulatory approval before being incorporated into radiology departments. The first FDA-approved CAD was a mammography-detection system, which localises breast lesions.⁸ Later, CAD was commercialised in other diseases such as lung cancer.  Some CAD diagnostic systems not only localise abnormal structures but also categorise them into subtypes indicative of certain diseases. Many new systems have been incorporated into radiology departments in recent years.

Current status of Ai-assisted medical imaging

What has been achieved by Ai assistants

Improved diagnostic accuracy

CAD can now provide second opinions to radiologists. Breast cancer detection CAD was shown to improve cancer detection rate meaning, more cancer cases, previously not perceived, could be identified.⁸ This helps reduce the chance of breast cancer progressing to a more advanced stage, which could otherwise be difficult to treat. For lung cancers, many CADs were designed to help distinguish between benign and malignant lung nodules. In some cases, radiologists assisted by CAD diagnostic systems perform better than radiologists alone.⁷ 

Improved operational efficiency

Radiology departments embedding AI into their workflow also enjoy higher efficiency. When assisted by CADs, radiologists can make faster diagnoses with fewer errors.⁹ This is especially important when the demand for imaging is high, such as during major trauma incidents and health emergencies like the COVID-19 pandemic.¹⁰ AI assistants have the potential to command other routine parts of the radiology workflow, such as scheduling appointments and drafting template reports, allowing time to focus on more complex clinical discussions.¹¹

Obstacles still present

In spite of recent success, further advancements still face obstacles. These include various ethical constraints, legal matters and practical issues.

Data privacy

One major concern for patients is data privacy. Personal medical information is highly confidential and it is completely understandable not to want your scans, along with your medical history, to be shared widely with AI companies. Although the transfer of certain data may be prohibited at a national level, there is concern about how this data could be shared more widely, across borders, where the legal and regulatory frameworks may differ.

Ai bias

Another ethical and practical issue is bias in AI. You may wonder how machines can be biased. The most straightforward answer is that steps in training the AI systems unavoidably include human biases that lead to an imbalance in data quality.⁶ The demographics of some populations may be underrepresented by the input data leading in turn to less accurate diagnoses. When these data are used to train AI models, biased judgments can be replicated. This can be further exacerbated if individuals not included in the original input algorithm also have reduced access to healthcare. 

Legal liability

An important legal consideration is who should be responsible for any errors. Although AI models have the potential to achieve very high accuracy, errors are inevitable. When mistakes are made that significantly affect a patient’s health, should the creators (AI companies), the users (radiologists) or other bodies (regulators or healthcare trusts) be held liable? In healthcare law, this is an issue that must be resolved before fully incorporating AI systems into radiology.⁶

The black-box problem

One of the practical problems of CADs is the lack of transparency. Advanced algorithms used in CAD typically cannot explain their reasoning. The lack of interpretability often leaves radiologists unable to trust the suggestions made by AI. This scepticism is compounded by the instinctive reluctance of many radiologists to fully adopt AI systems.¹²

The future of Ai-assisted medical imaging

You can now see how valuable AI tools could be in medical imaging. According to four scientists at McKinsey & Company, “It is hard to imagine that AI will not ultimately transform radiology.”¹² As various regulatory and legal hurdles are overcome, this transformation is closer to becoming a reality. As people get more familiar with protocols for recording good-quality data, and advanced healthcare becomes more accessible to underrepresented groups, AI bias will present a less significant problem. Further, as the application of AI in other fields matures, relevant regulations will be developed which can be translated to practice in radiology. 

The incorporation of AI into radiology is expected to result in speedier reporting and improved accuracy of radiology reports. 

Summary

Like many other industries, medical imaging has been significantly shaken by the emergence of AI. Many invaluable CAD systems have been developed by advancements in computing power, ML algorithms, and the increasing access to radiology databases. If more efforts are made to surmount legal, ethical, and practical constraints, the wider application of AI-assisted medical imaging will yield major benefits for radiologists and patients in the future.