AI in Radiology

Introduction

Self-driving cars used to live within science fiction stories, but now, with the help of artificial intelligence (AI), they are progressing towards becoming a reality. This is just one example of how AI has the potential to permeate our everyday lives. From self-driving cars using machine learning technology to object recognition using deep learning technology, AI has demonstrated its prowess and is revolutionising our lives.

And yet, this is only the beginning.

One of the most profound impacts of AI is unfolding in healthcare, where its integration has changed how we monitor, diagnose and treat diseases. Complementing human perception, AI can discover and learn relationships within multimodal clinical data that might otherwise go unnoticed.

This article will explore how AI intersects with radiology and its part in advancing diagnostic accuracy, streamlining workflows, and increasing support for clinical-based decision-making, thereby improving patient outcomes and experience. We will also discuss the challenges of using AI in healthcare and the future direction this technology might take us.

Background information

Understanding AI

AI enables machines and computers to perform advanced functions such as simulating human intelligence and facilitating problem-solving. By processing large datasets (also known as big data), AI analyses and identifies relationships and correlations and subsequently uses these insights to make predictions.

The main process of AI is machine learning, where algorithms can learn from data and improve their future processes over time. Deep learning, a subset of machine learning, uses an artificial neural network to mimic the human brain’s operations.

At its core, AI is used for data analytics, predictions, forecasting and object categorisation, among many other applications.

Understanding radiology

Radiology uses medical imaging examinations to capture images of the internal body that are not visible to the human eye. By acquiring these images of the internal body structure, healthcare practitioners can visualise the presence or absence of any health anomalies, acting as a diagnostic and treatment planning aid.

Typical radiology workflow

This is how a typical radiology workflow would operate:

1. Image acquisition
   Medical imaging modalities used to acquire images include, but are not limited, to: X-rays, computed tomography (CT) scans, and magnetic resonance imaging (MRIs)
2. Image and data interpretation
   Healthcare practitioners then process the medical images and interpret them based on their expertise
3. Outcomes and clinical action
   Using the insights gathered, healthcare practitioners determine the condition of the patients. Alongside other multimodal clinical data, practitioners from various fields collaborate to craft a clinical treatment plan for the patient

Importance of radiology

Radiology is essential for detecting, assessing, and monitoring various conditions – from fractures to heart conditions. It plays a vital role in determining appropriate treatment methods that can be used to improve patients' outcomes. Based on data from the United Nations Scientific Committee on the Effects of Atomic Radiation (UNSCEAR), 4.2 billion medical imaging exams were performed worldwide between 2009 and 2018.¹

Medical imaging examinations were issued at 12.5% of patient visits.² Given the benefits radiology brings to the diagnosis and monitoring process, its increasing importance in the clinical workflow process can be seen from the surge in demand for radiological examinations.³ 

Challenges faced within radiology

Reliance on expertise

Interpretation of medical images relies largely on the healthcare practitioner’s skill and experience. Without proper training and experience, practitioners may even struggle to capture a clear image at the correct position, which is crucial to getting an accurate visual image. Adequate understanding, skillful visual identification, and assessment of abnormalities are essential, and these skills vary significantly among practitioners.

Prone to human error — cognitive error and perceptual error

Radiology involves diagnosis and decision-making under uncertainty, making it susceptible to human error. This perception and interpretation process is subjective and can vary between individual practitioners, potentially leading to missed or incorrect diagnoses, which impacts patient outcomes.

Cognitive errors occur when an abnormality is identified but healthcare practitioners misinterpret its significance. Studies have shown that 20–40% of the total errors are cognitive errors.⁴ 

Perceptual errors, on the other hand, are errors in which practitioners fail to recognise the abnormality from the get-go due to oversight. A possible reason may be the large amount of imaging data they analyse daily. This type of error accounts for 60–80% of errors.⁴

Labour intensive and time consuming

The identification and analysis processes in radiology such as volume measurements and radiograph interpretation are labour-intensive and time-consuming. Radiology involves capturing and interpreting medical images under time constraints, with the need to correlate complex multimodal clinical data adding to the workload.

Intersection between AI and radiology

One of the main ways in which AI is incorporated in radiology is radiomics. Instead of medical digital images being qualitatively assessed and interpreted by healthcare practitioners, they are extracted and converted into quantitative data to give practitioners a more accurate understanding of the disease.

These mineable radiomics data are extracted from a large number of quantitative features like size, texture, and shape. These data act as ‘big data’, used to create a database and mined further to train machine-learning systems to recognise key features from the data or predict outcomes. The introduction of radiomics allows the recognition of relationships in imaging data and provides a quantitative assessment of features.

AI can be used in every aspect of radiology such as:

• Automating image segmentation, lesion detection, measurement, labelling and comparing with past images
• Generating and detecting errors in radiology reports
• Data mining in research
• Workflow optimisation and improvement in outcome measures

Benefits of AI in radiology

Diagnostic accuracy

• Radiomic information extraction
  • These radiomic quantitative data are more extensive and comprehensive, extracting data not visible to the naked eye. Processes using advanced technological systems have shown increased medical accuracy compared to those without⁵ 

Efficiency and efficacy in healthcare workflow

• Automation of labour-intensive and time-consuming tasks
  • AI can automate tasks like image segmentation lesion detection and volume measurements, allowing healthcare practitioners to free up time to focus on more complex cases
• Case prioritisation
  • AI surveillance programs can identify suspicious or positive cases for early review while expediting clear and direct cases
• Standardisation
  • AI can create a basis of standardisation, reducing variability in diagnostic outcomes and enhancing the reliability of treatment plans. This allows for greater clarity and transparency in the diagnosis and treatment process
• Faster turnaround time
  • Patients can receive their results even faster, decreasing their waiting time, and enhancing the patient’s overall experience and satisfaction

AI-based support for clinical decision-making

• Optimal clinical decisions
  • Healthcare practitioners can better acquire and comprehensively analyse patients’ data through the integration of radiomic data with other multimodal clinical data like genomic data and pathology reports to derive optimal clinical decisions⁶
• Reduced errors
  • AI-based clinical decision support systems can improve the accuracy of diagnoses and reduce errors, thereby enhancing the reliability of clinical decision-making
• Enhanced patient care
  • AI complements healthcare practitioners’ analysis, ensuring accuracy and improving overall patient care

Real-world application

COVID-19

Using AI, healthcare practitioners were able to detect changes in lung features associated with COVID-19. AI systems incorporating deep learning software to computer-aided detection (CAD) were used to deduce both the presence of the virus infection and the severity of the disease.

Through medical images, the CAD system sieves through data like the opacity pattern and volume inside the lungs and deduces COVID-19 patients and the severity of their cases.⁷ Using this information, doctors can more accurately diagnose and treat COVID-19 patients.

Challenges of incorporating AI in radiology

Data quality

AI systems rely on high-quality data for training and development. The performance of AI models depends significantly on the quality and comprehensiveness of the datasets they are trained on. Given the large datasets, many correlations can be easily deciphered but correlation does not mean causality.⁶ Causality is harder to establish despite the use of massive datasets.

Bias and fairness

Systemic bias can be introduced throughout the design, development, and testing process. Within the training data, selection bias from demographics of the training data can also be amplified leading to disparities in healthcare outcomes. The findings from the training data may not be transferable to other target populations.

The curation of representative datasets is essential and developers of the AI system have to use robust and diverse datasets for training.

Governance

Effective governance frameworks and clear standards are essential for the responsible development and incorporation of AI in radiology. These frameworks should address issues related to the implementation and development of systems, data privacy, security and regulatory compliance.

Robust data governance policies must be established to protect patients’ data from breaches and exploitation. Ensuring the ethical use of data collected is needed to protect sensitive health information.

Liability

When faced with situations where AI systems make errors or patients experience adverse outcomes, liability and accountability become an issue. It can be difficult to determine who is liable — the healthcare provider who used the system or the AI developer in charge of the system.

Clear guidelines and frameworks are needed to assign responsibility appropriately.

Future directions

Continuous research, collaboration and education

Continuing extensive research is vital for developing and refining AI algorithms that reflect the target population. Strong collaborations between healthcare practitioners and AI developers ensure that AI solutions are tailored to meet clinical needs. AI developers must be educated on the medical aspect of the data to tailor a system that reflects the process, and healthcare practitioners must learn how to use AI systems within their operations.

Incorporation of AI in other medical fields

Dermatology

Machine learning systems can analyse skin lesion images, and identify patterns and features indicative of skin diseases, potentially improving early detection of diseases like melanoma.

However, for this implementation to be successful, the AI must be trained on datasets that feature a diverse array of skin conditions on different skin colours. This is because different skin conditions may appear differently on different skin colours.

Pathology

Machine learning systems can analyse digital pathology slides, identify abnormalities in tissue samples, and help pathologists detect diseases like cancer more accurately and efficiently.

Summary

AI in radiology is a rapidly developing field with significant benefits to reap. AI would complement the existing process by increasing accuracy, efficiency, and support in the clinical process. Continuous research, education, and collaboration between AI developers and healthcare practitioners will be crucial in harnessing the full potential of AI in this field, from improving patient care and outcomes in radiology to other medical disciplines such as dermatology and pathology.