A study by the Centre for Advanced Research in Imaging, Neuroscience, and Genomics (CARING) recently found that AI can help doctors of radiology provide more accurate, actionable reports to make better treatment decisions. AI can also help physicians avoid unwarranted procedures that are the result of ambiguous reports.
Many forms of bias can negatively affect a radiologist’s analysis. For instance, availability bias may cause a radiologist to associate similar images they have seen recently with a current case. Similarly, alliterative bias allows the misdiagnosis of one radiologist to affect another’s judgment. Regret bias may cause a radiologist to overestimate the likelihood of disease because of possible outcomes that could result from failing to diagnose the disease.
CARING Compares Human Radiologists to AI
CARING initially started its study to find evidence of hedging in radiology reports, and also hoped to assess the accuracy and validity of AI-generated radiology reports. The organization postulated that even when doctors of radiology are confident there is no disease, they make vague statements to protect themselves from errors and lawsuits. In comparison, properly trained AI has no such concern and generates clearer, unbiased reports.
The study showed AI could help reassure a radiologist of their opinion and give them the confidence they need to express it. It involved the use of 297 chest X-ray images and an AI that detected abnormalities in the X-rays and autogenerated a clinical report. The models used to train the AI algorithms were based on nearly 1 million chest X-rays from various sources.
AI was used to analyze the medical imaging data and generate a preliminary radiology report. The AI-generated report was then compared to a report prepared by a radiologist. The study found that nearly 80% of the AI preliminary reports were clinically accurate. In 5% of cases, they were even more accurate than the radiologist’s report.
AI Is a Valuable Tool to Help Radiologists
Researchers say the AI reports and the radiologist reports would likely never agree 100%. Even radiologists often have differing opinions on an X-ray report. Nevertheless, an important distinction to make is that AI is intended to help radiologists make diagnoses, not replace them. AI algorithms scan through images carefully and look for patterns and discrepancies that a doctor might miss, resulting in fewer misdiagnoses or complications.
Healthcare professionals recognize that AI can help improve the efficiency and accuracy of radiology, but some question whether a machine should be trusted for the diagnosis of human disease. They fear that human bias could work its way into machine learning models. Fortunately, the CARING study seems to support AI analysis as a viable option.