Medical Imaging

     In 2016, Geoffrey Hinton, a prominent AI researcher, said we should stop training radiologists. He claimed it was “completely obvious” within five years; AI deep learning would outperform radiologists. There has been significant hype around AI in radiology, with more than a few predicting AI would soon replace radiologists or diagnose diseases better.

Despite the bold predictions, the reality has been quite different. Seven years after Hinton’s prediction, not a single radiologist had been replaced by autonomous software. In fact, there is still a high demand for radiologists, with many facilities struggling to fill positions. Most experts now view AI as complementary to radiologists.

The consensus is AI will enhance and augment radiologists’ capabilities. AI is expected to improve workflow efficiency, allowing radiologists to focus on more complex cases. Recent research suggests AI’s benefits vary among individual radiologists, emphasizing the need for personalized integration of AI tools.

As a side note, I believe personalization of AI is the next big thing in AI. chatGPT and the numerous other chatbots are the beginning for AI. I believe what comes next will be AI agents which are specific to industry segments such as healthcare, insurance, real estate, manufacturing, retail, etc. More on this in a future blog post.

I believe the bigger issue than replacing radiologists is enabling radiologists to be more accurate. A recent study published in Radiology estimates approximately 4.2 billion medical imaging exams are performed globally each year. Even under optimal working conditions, the error rate for plain film radiography is estimated to be around 3-5%. For cross-sectional imaging like CT and MRI, error rates are significantly higher, ranging from 20-30%. Not all errors lead to clinically significant consequences. The types of errors can vary from minor discrepancies to major misinterpretations which could affect patient care.

The implications of these error rates are significant. If we assume an average 4% error rate across all imaging modalities, with 4.2 billion exams, the number of potential errors could reach 168 million per year. Fortunately, efforts are ongoing to reduce error rates through various strategies, including AI assistance, double reading, peer review, and improved training and awareness among radiologists.

Based on AI search results, the most common types of medical imaging errors fall in a number of categories. Perceptual errors are the most prevalent, accounting for 60-80% of all radiology errors. These occur when radiologists fail to detect or recognize abnormalities present in the images. The main types of perceptual errors include Missed Findings (underreading), which are the most common errors, occurring in about 42% of cases. This type of error happens when a radiologist fails to identify an abnormality present on the image.

A second type of error is called Satisfaction of Search (SOS). This second most common error occurs in about 22% of cases. This happens when a radiologist stops searching for additional abnormalities after finding one, potentially missing other important findings. A third type of perceptual error is called Faulty Reasoning. These occur in about 9% of cases. Such errors involve correctly identifying an abnormality but attributing it to the wrong cause.

Cognitive Errors occur during the interpretation phase and include two types. First is Misinterpretation. This involves incorrectly interpreting an abnormal finding or, rarely, misinterpreting a normal finding as abnormal. Hopefully not frequent but some errors are a result of a Lack of Knowledge. These errors result from insufficient expertise or experience in identifying certain conditions.

Technical Errors are related to the imaging process itself. First is Poor Image Quality. This can result from incorrect patient positioning, equipment malfunction, or improper imaging techniques. I have experienced this myself. Hopefully rare, but another source of errors is Inappropriate Selection of Imaging Procedures. These errors result from using the wrong type of imaging study for a particular clinical question.

Also, probably rare but errors sometimes are a result of Failure to Communicate Findings. This involves not effectively conveying important results to referring physicians or patients. Incomplete Reporting can happen with complex results, omitting crucial information or recommendations in the radiology reports.

 Those are the broad spectrum of possible errors. Some more specific areas of concern would include Missed Fractures. These are especially common in emergency settings, especially in areas like the femur, navicular bone, and cervical spine. Missed Cancer Diagnoses commonly occur in lung nodules on chest radiographs, breast lesions on mammograms, and colorectal carcinomas on barium enema studies.

A final type of error is Overlooked Findings Outside the Primary Area of Interest. In other words, missing abnormalities visible but outside the main region being examined. By understanding these common error types, radiologists and healthcare systems can implement strategies to reduce their occurrence and improve patient care. AI will play a major role in this.

I believe benefits of AI in radiology will be significant. AI enhances diagnostic precision by analyzing medical images with remarkable accuracy, often surpassing human capabilities in detecting abnormalities like tumors or fractures. Studies report AI-assisted radiologists achieve consistently higher accuracy in diagnosis, reducing errors and improving patient outcomes. AI can help optimize workflow by automating routine tasks such as image sorting and preliminary assessments, allowing radiologists to focus on complex cases. AI can significantly reduce turnaround times for diagnostic reports. One study showed a reduction from 11.2 days to 2.7 days.

Reduced workload and burnout can be achieved by using AI to handle repetitive tasks. The result can be alleviated stress and workload for radiologists, helping them maintain high levels of care and accuracy. AI supports personalized medicine by integrating predictive analytics and correlating imaging data with treatment plans. Applications such as dose reduction in imaging and faster scanning procedures improve efficiency while maintaining quality.

The potential benefits are huge but there are challenges and limitations. There is a risk of radiologists relying too heavily on AI recommendations, which could lead to errors if the AI system is inaccurate. The impact from AI may not be consistent. The benefits of AI vary among radiologists based on factors like experience and familiarity with the technology. In some cases, inaccurate AI predictions can negatively affect performance.

AI is only as good as the data it uses. I believe this issue will get increased attention. Equal to that is the “black box” nature of AI algorithms. As I have written before, clinicians have been skeptical of AI diagnoses and treatment plans when they don’t know exactly how the AI algorithms reached its conclusions. Some hospital networks may not have a robust infrastructure to support AI implementation. Despite its promise, the clinical adoption of AI remains limited due to gaps between proof-of-concept studies and real-world applications.

What is the future? Experts believe while AI will not replace radiologists, it will redefine their roles by acting as a powerful tool for improving diagnostic capabilities and workflow efficiency. Early adopters of AI are expected to lead advancements in the field, but successful integration requires addressing current limitations through personalized approaches and ongoing monitoring of AI performance in clinical settings. Based on my readings and the course I have been taking from MIT, I am optimistic the 4% error rate will be reduced significantly and the overall process improvements will lead to better patient outcomes and hospital efficiency.

Perplexity Pro was my research partner for this post. If you are interested in reviewing the sources backing up the various numbers and assertions, the sources of information for the post are here.

Note: I use Perplexity, ChatGPT, and Gemini AI chatbots as my research assistants. AI can boost productivity for anyone who creates content. Sometimes I get incorrect data from AI, and when something looks suspicious, I dig deeper. Sometimes the data varies by sources where AI finds it. I take responsibility for my posts and if anyone spots an error, I will appreciate knowing it, and will correct it.