The State of AI in Radiology

Imagine yourself as a doctor, tasked with interpreting medical scans. When someone breaks their arm, has surgery, or symptoms of a terminal disease, you are the hero who swoops in and gives a diagnosis. This is the job of radiologists, physicians who read X-rays, CT scans, MRIs, and ultrasounds, which are all scans that doctors order to visualize what is happening inside the human body. These medical images become a crucial part of the portfolio that doctors build to understand what might be wrong with a patient.

When you seek medical attention at a hospital, you place the fate of your physical health in the hands of a radiologist. You need to strongly trust in their ability to diagnose your condition and recommend a proper treatment. Consequently, this makes it profoundly upsetting if the radiologist were to misdiagnose you. Shockingly, such occurrences are not uncommon–in fact, 71% of radiologists reported that they have been named in malpractice suits (Stempniak, 2022). The majority of these legal actions are reported to be caused by missed diagnoses, often attributed to human error by many in the medical field.

To help minimize these problems, many argue that we should integrate artificial intelligence into the diagnostic process of radiologists, such as using machine learning models to prescreen patient scans (Hosny et al., 2018). While these new tools certainly show promise, and are currently a field of active research and development, we should be measured in adopting them in the healthcare industry. Rushing to implement AI solutions without thorough evaluation and consideration of their potential consequences means that we risk undermining the very essence of compassionate and patient-centric healthcare.

Compassionate, patient-centric healthcare is vital in the medical field. It recognizes the humanity of patients, fostering trust, comfort, and better health outcomes. In an era where machine learning is advancing rapidly in medicine, maintaining this focus becomes even more crucial. Machine learning can provide personalized treatment plans, predict patient needs, and streamline administrative tasks, allowing healthcare providers to spend more time with patients. However, it could also undermine this goal by depersonalizing interactions, relying solely on data without considering individual circumstances, and perpetuating biases present in data. So while machine learning offers promising opportunities, it must be carefully integrated into healthcare systems to ensure that patient wellbeing remains at the forefront of medical practice.

An X-rayed Gorilla

While radiologists play a crucial role in diagnosing patients, it is evident that they are not infallible, as shown by a 2013 study on inattentional blindness in radiologists. Inattentional blindness occurs when someone is paying attention to something, but they fail to see a stimulus despite its being in plain sight. Researchers from Harvard Medical School and Brigham and Women’s Hospital in Boston demonstrated inattentional blindness in radiologists through a striking experiment (Drew et al., 2013). They showed the radiologists typical CT scans of the lung, and asked them to identify if the patients had lung nodules, which is a very common diagnosis.

However, these scans had one crucial difference from ones imaged from a normal patient. The scientists added a small gorilla, about the size of a matchbox, in the upper right hand corner of the scan. Reassuringly, the majority of radiologists successfully identified the nodules in the lungs. But 83% of them did not even notice the gorilla, despite a majority of them looking directly at the gorilla when it was visible on the scan (Drew et al., 2013). None of the untrained participants in the study noted the gorilla, either. Interestingly, the gorilla was 48 times the size of the lung nodules they found. All of the subjects, both radiologists and not, saw the gorilla after being asked if they noticed anything unusual with the image. Below is a figure from the experiment showing the images presented to the radiologists, as well as their eye movements while interpreting the scan.

This radiologist clearly looked at the gorilla, as the figure shows that their eyes focused on it during the experiment. However, they did not process the gorilla’s presence. This study is fundamental to understanding the human limitations doctors face in treating patients. The nature of the human brain makes it such that radiologists are not perfect at their jobs and may not always notice every abnormality.

It is important to note that these findings–and the fact that radiologists are not perfect–do not suggest that radiologists are terrible at their jobs. It just means that despite being highly specialized, highly trained individuals, they are only human. Their job is quite demanding and has high stakes. Aside from inattentional blindness, Drew et al. (2013) also posit that radiologists could also be suffering from the effect of a phenomenon known as “satisfaction of search,” or SoS. This cognitive bias occurs when finding one stimulus makes the searcher experience a sense of accomplishment. In this case, the radiologist’s satisfaction with finding the first abnormality leads them to overlook other ones present in the same image.

AI Enters The Playing Field

It is tempting to believe that artificial intelligence provides the perfect solution to these problems. Proponents of using artificial intelligence in the field of medical imaging argue that computers are not as fallible as humans, because it is often a lapse in awareness that causes devastating misdiagnoses (Hults et al., 2024). Properly implemented artificial intelligence models, that do not lose focus, can make diagnostic radiology more accurate and efficient. Machines will not make the same mistakes as humans. They never get tired, hungry, or frustrated, nor can they fall prey to cognitive problems such as inattentional blindness or SoS. However, as we will discuss later, they have other, potentially more worrisome pitfalls.

Furthermore, the demand for interpretation of X-rays, CT scans, and MRIs has skyrocketed as technology develops and we gain the ability to better see inside the human body. But the supply of radiologists to read these scans has not kept pace. In fact, some reports show that to meet the demand in a typical workday, the average radiologist needs to read one image every few seconds (Hosny et al., 2018), amplifying the risk of incorrect diagnosis and patient harm.

Area of Active Research

This pressing challenge has prompted companies all over the world to actively explore AI solutions for clinicians. Google, for example, is researching AI techniques to enhance the diagnosis and treatment of various cancers, such as skin, eye, lung, and breast cancer (AI Imaging & Diagnostics – Google Health, n.d.). Such endeavors underscore a broader trend of artificial intelligence permeating various sectors, from Apple’s ventures into developing a self-driving car to companies using AI to help farmers increase crop yields (Partridge et al., 2023). Indeed, artificial intelligence is truly affecting and reshaping every facet of society.

As we watch the successes and failures of the application of artificial intelligence to other industries, people wonder how it will fare in the healthcare industry. It seems that some industry giants want to dive headlong into applying machine learning to as many fields as they possibly can. With advanced imaging technology and artificial intelligence, brain tumors can be detected three times faster than by a human (Artificial Intelligence Expedites Brain Tumor Diagnosis, 2020). While that exploration is a worthwhile pursuit, as a society we cannot so cavalierly use these tools in the healthcare industry.

The domain of healthcare stands far apart from driving, agriculture, or manufacturing, to name a few fields. Health is an incredibly delicate and personal aspect of our lives, a realm where every minute detail and decision directly impacts individual well-being. While the promise of artificial intelligence in revolutionizing medical imaging, and by extension healthcare, is undeniably tantalizing, we must tread cautiously. Every algorithmic recommendation, each automated diagnosis, carries profound implications for patient care, trust, and autonomy.

What About Healthcare Inequity?

To produce the impressive results that machine learning models do, they require a large amount of data to train on. In healthcare, this information is difficult to acquire. Much of the data is strongly protected by patient privacy laws, making the collection of large data sets a challenging task (Tang, 2020). Imagine you are planning an elaborate gala. In order to throw a bash, you need a guest list–you have to collect your guests’ names, addresses, phone numbers, and dietary preferences. This information becomes your dataset. In medicine, these datasets might include information about surgeries patients have had, or treatments they have undergone. Understandably, one might not want the entire world knowing this information.

On top of that, AI systems often inadvertently perpetuate and amplify biases in the data used to train them. Using artificial intelligence to decide patient outcomes could lead to disparities in diagnosis and treatment among demographic groups, exacerbating health inequities (McMains, 2023). For example, demographics that are overrepresented in datasets–in other words people who already have access to healthcare–might get more accurate diagnoses compared to those who already struggle to access it. For example, racial and ethnic minorities in America have much higher chronic disease and early death rates than the racial majority, and indigenous populations such as Native Americans and Alaskan Indians have an infant mortality rate that is 60 percent higher than the rate for their white counterparts (Baciu et al., 2017), in part due to unequal healthcare access rates. AI works by training on this biased data, creating mathematical trends from it, and using these trends to make predictions. So, since the data is biased, the trends become biased, and then the AI makes biased predictions. These predictions can make healthcare access worse for people who already don’t have easy access, or it can even cause a misdiagnosis that harms a patient. Before we use these tools in practice, we must analyze, prepare for, and address these issues.

Furthermore, artificial intelligence currently operates as a “black box,” meaning their decision making process is not transparent to human users (Monardo 2023). Some examples of black boxes include your mobile phone or car. You rely on the functionality of these items without understanding the intricate systems that allow them to work. But a black box is incredibly dangerous for healthcare. It is deeply important that patients exercise autonomy and control over their own healthcare plans. Using AI might limit opportunities for shared decision making between the patient and their healthcare provider, particularly if the patient feels pressured to blindly follow algorithmic recommendations without fully understanding the implications. The algorithm might not even be trained to make the best recommendation for the patient, if they are not in an affluent, overrepresented demographic.

What’s Really Happening Underneath the Surface?

From the perspective of the healthcare provider, the reliance on AI in medical imaging also poses unique challenges due to the lack of transparency. Unlike consulting with a human expert, there is no opportunity to directly question or clarify how the AI arrived at a particular diagnosis because of its black box nature. This opacity introduces a significant risk of missed findings or misinterpretations that may go unnoticed by expert radiologists. The inherent complexity of medical images means that subtle nuances might be overlooked by the AI, leading to potential diagnostic errors. So if AI experiences issues like SoS earlier, the radiologist will not be able to tell. Healthcare providers must navigate a delicate balance between the efficiency of AI-assisted diagnosis while remaining aware of the danger of bad recommendations. Using the AI requires them to place blind trust in its capabilities, with limited to no insight into the underlying rationale behind its recommendations.

Lastly, there are serious ethical considerations regarding patient consent and data privacy. As AI algorithms need to analyze vast swaths of data to generate their models (Monardo, 2023), questions arise surrounding patients’ awareness of and consent to the use of their personal health data. Additionally, the issue of keeping all of this sensitive patient data secure is a grave one. As we increasingly live our lives online, the number of large scale data breaches has drastically risen. Breaches represent a clear privacy violation for patients, exposing their personal data and potentially causing embarrassment or discrimination. They could also lead to medical identity theft, where criminals steal their data to get medical services, prescription drugs, or insurance coverage fraudulently. Implementing AI while also adequately addressing data security and privacy concerns is an incredible challenge (Tang, 2020). Yet, it is critical to upholding patient trust and confidentiality.

Striking a Delicate Balance

At its core, the incorporation of AI into medical imaging represents a pivotal juncture for us all. We must balance automation and human judgment. On one hand, AI offers unparalleled capabilities in the analysis of complex medical images, providing fast and potentially more accurate medical diagnoses than human methods. However, the reliance on AI also raises fundamental questions. While AI excels at processing vast amounts of data and identifying visual patterns, it may lack the nuanced understanding inherent to human cognition. Healthcare professionals must learn to strike the right balance between deferring to AI and exercising their own clinical expertise. There is a natural tension between AI-driven automation and the more empathetic human approach.

As we navigate this transformative era, it is imperative that we prioritize not just technological advancement, but also ethical considerations, patient-focused care, and the trust in a doctor-patient relationship. It is undeniable that artificial intelligence will continue to bring significant change to our society. Regardless of the nature of this change, our approach to integrating AI into medical imaging must be guided by empathy, responsibility, and a steadfast commitment to upholding the highest standards of healthcare.

 

References

AI Imaging & Diagnostics – Google Health. (n.d.). <https://health.google/health-research/imaging-and-diagnostics/>

Artificial intelligence expedites brain tumor diagnosis. (2020, February 12). Cancer.gov. <https://www.cancer.gov/news-events/cancer-currents-blog/2020/artificial-intelligence-brain-tumor-diagnosis-surgery>.

Baciu, A., Negussie, Y., & Geller, A. (2017). The state of health disparities in the United States. Communities in Action: Pathways to Health Equity. <https://www.ncbi.nlm.nih.gov/books/NBK425844/>

Drew, T., Võ, M. L., & Wolfe, J. M. (2013). The Invisible Gorilla strikes again. Psychological Science, 24(9), 1848–1853. <https://doi.org/10.1177/0956797613479386>

Hosny, A., Parmar, C., Quackenbush, J., Schwartz, L. H., & Aerts, H. J. W. L. (2018). Artificial intelligence in radiology. Nature Reviews. Cancer, 18(8), 500–510. <https://doi.org/10.1038/s41568-018-0016-5>

Hults, C. M., Ding, Y., Xie, G. G., Raja, R., Johnson, W., Lee, A., & Simons, D. J. (2024). Inattentional blindness in medicine. Cognitive Research Principles and Implications, 9(1). <https://doi.org/10.1186/s41235-024-00537-x>

McMains, V. (2023, May 2). 2023 News – AI in medical imaging could magnify health inequities, study finds | University of Maryland School of Medicine. <https://www.medschool.umaryland.edu/news/2023/ai-in-medical-imaging-could-magnify-health-inequit%20ies-study-finds.html>

Monardo, V. (2023, September 6). 6.3900: Lecture 1 material. Cambridge, MA.

Partridge, J., Inman, P., Lawson, A., Jolly, J., Partington, R., Topham, G., Makortoff, K., & Butler, S. (2023, February 18). Technology | The Guardian. <https://www.theguardian.com/technology/2023/feb/18/from-retail-to-transport-how-ai-is-changing-ever%20y-corner-of-the-economy>

Stempniak, M. (2022, February 11). 71% of radiologists named in a malpractice lawsuit, though numbers are down during pandemic. Radiology Business. <https://radiologybusiness.com/topics/healthcare-policy/radiologists-malpractice-lawsuit-trends-medscape>

Tang, X. (2020). The role of artificial intelligence in medical imaging research. BJR|Open, 2(1), 20190031. <https://doi.org/10.1259/bjro.20190031>

 

Return to Table of Contents