GPT-4o's 'all or nothing' accuracy continues to hinder its radiologic capabilities
OpenAI’s generative pre-trained transformer ChatGPT has come a long way since its inception, now with the ability to process images, new data suggest it still needs extensive training before it can be trusted in medical settings.
Published in Cureus, the new data detail GPT-4o's ability to interpret medical imaging. GPT-4o has been trained to process and generate text, images and audio, which has made it a target for researchers seeking ways to improve both radiology workflows and access to quality care in rural settings.
“While both its text-based interpretation and image recognition skills are remarkable, the latter has been particularly impactful in areas such as diagnosing cancer on histopathology, differentiating between skin lesions, or assessing lung injury,” corresponding author Nikoloz Papiashvili, a medical student at Tbilisi State Medical University in Tbilisi, Georgia, and co-authors noted. “This impact has led to a focus on the field, further amplified by the shortage of qualified radiologists in underserved communities and the technology’s ability to enable faster decision-making in emergency settings.”
Its potential is evident; however, the team’s findings indicate that the performance of GPT-4o depends largely on context and comparison when it comes to reading medical images.
For their study, researchers presented GPT-4o with 377 cases that included imaging—X-ray, CT and MRI—across various organ systems. A standard prompt was used, and the model was not given clinical context nor prior imaging to analyze. Three radiologists rated the responses on a 5-point scale.
The model performed better when reading X-ray images than it did with CTs, accurately interpreting twice as many radiographs. It was six times more likely to provide accurate diagnoses with images of the abdomen compared to the pelvis, and nearly three times more accurate in identifying bleeding conditions in comparison to neoplastic findings.
“The overall distribution of median ratings demonstrates a bimodal distribution, indicating that most cases were interpreted either completely incorrectly or entirely without error, with only a small percentage of responses being partially accurate,” the authors explained “This finding suggests that the AI adopts an ‘all or nothing’ approach, expertly recognizing anything it has seen before but being practically unable to work with anything it has not visually encountered, despite having access to theoretical knowledge on these conditions.”
One factor affecting GPT’s accuracy was hallucinations, the group noted. This is a known issue with ChatGPT that causes it to present false information as factual. In other words, when GPT provided an inaccurate interpretation, it did so with conviction, which could be problematic for users without medical expertise who take the information it provides as truth. What’s more, if this technology were to be deployed in medical settings as is, it could be misleading for trainees or less experienced providers who are seeking information on a topic they are not well informed about.
From their observations on its performance, the authors offered several suggestions for future studies on GPT-4o's medical capabilities.
“Future research should focus on methods to combat these pitfalls. One possibility may be to instruct the AI to analyze the image in sections, with each part contributing to a consolidated list of potential diagnoses. This approach has shown particular success in the field of pathology,” the group advised. “Additionally, text-based data could be incorporated as a secondary source of information, supporting or challenging certain diagnoses that the model has already considered.”
Read more about their observations and suggestions here.
In addition to her background in journalism, Hannah also has patient-facing experience in clinical settings, having spent more than 12 years working as a registered rad tech. She began covering the medical imaging industry for Innovate Healthcare in 2021.