Summary: ChatGPT-4 Vision shows strong performance on text-based radiology questions but struggles with image interpretation, highlighting limitations in its current application to radiology.

Key Takeaways

  • Mixed Performance: ChatGPT-4 Vision performed well on text-based radiology questions but struggled significantly with image-based ones, highlighting limitations in its current ability to interpret radiologic images accurately.
  • Promising Applications: Despite its challenges, GPT-4 Vision shows potential for assisting radiologists with tasks such as simplifying radiology reports and selecting appropriate imaging protocols.
  • Need for Improvement: The study emphasizes the need for more specialized evaluation methods and improvements in image interpretation to enhance the model’s reliability in critical radiology tasks.

———————————————————————————————————————————————————

Researchers found that ChatGPT-4 Vision excelled at answering text-based radiology questions but struggled with image-related ones. The study, published in Radiology by the Radiological Society of North America (RSNA), highlights the model’s mixed performance. ChatGPT-4 Vision is the first version of the model capable of interpreting both text and images.

ChatGPT-4 Vision’s Mixed Radiology Performance

“ChatGPT-4 has shown promise for assisting radiologists in tasks such as simplifying patient-facing radiology reports and identifying the appropriate protocol for imaging exams,” says Chad Klochko, MD, musculoskeletal radiologist and artificial intelligence (AI) researcher at Henry Ford Health in Detroit. “With image processing capabilities, GPT-4 Vision allows for new potential applications in radiology.”

For the study, Klochko’s research team used retired questions from the American College of Radiology’s Diagnostic Radiology In-Training Examinations, a series of tests used to benchmark the progress of radiology residents. After excluding duplicates, the researchers used 377 questions across 13 domains, including 195 questions that were text-only and 182 that contained an image. 

GPT-4 Vision answered 246 of the 377 questions correctly, achieving an overall score of 65.3%. The model correctly answered 81.5% (159) of the 195 text-only queries and 47.8% (87) of the 182 questions with images.

“The 81.5% accuracy for text-only questions mirrors the performance of the model’s predecessor,” he says. “This consistency on text-based questions may suggest that the model has a degree of textual understanding in radiology.”

Genitourinary radiology was the only subspecialty for which GPT-4 Vision performed better on questions with images (67%, or 10 of 15) than text-only questions (57%, or 4 of 7). The model performed better on text-only questions in all other subspecialties.

The model performed best on image-based questions in the chest and genitourinary subspecialties, correctly answering 69% and 67% of the image-containing questions, respectively. The model performed lowest on image-containing questions in the nuclear medicine domain, correctly answering only 2 of 10 questions.

Prompt Impact on GPT-4 Vision

The study also evaluated the impact of various prompts on the performance of GPT-4 Vision.

  • Original: You are taking a radiology board exam. Images of the questions will be uploaded. Choose the correct answer for each question. 
  • Basic: Choose the single best answer in the following retired radiology board exam question. 
  • Short instruction: This is a retired radiology board exam question to gauge your medical knowledge. Choose the single best answer letter and do not provide any reasoning for your answer. 
  • Long instruction: You are a board-certified diagnostic radiologist taking an examination. Evaluate each question carefully and if the question additionally contains an image, please evaluate the image carefully to answer the question. Your response must include a single best answer choice. Failure to provide an answer choice will count as incorrect. 
  • Chain of thought: You are taking a retired board exam for research purposes. Given the provided image, think step by step for the provided question. 

Although the model correctly answered 183 of 265 questions with a basic prompt, it declined to answer 120 questions, most of which contained an image. “The phenomenon of declining to answer questions was something we hadn’t seen in our initial exploration of the model,” Klochko says. 

GPT-4 Vision Skips Many Image Questions

The short instruction prompt yielded the lowest accuracy (62.6%). On text-based questions, chain-of-thought prompting outperformed long instruction by 6.1%, basic by 6.8%, and original prompting style by 8.9%. There was no evidence to suggest performance differences between any two prompts on image-based questions.

“Our study showed evidence of hallucinatory responses when interpreting image findings,” Klochko says. “We noted an alarming tendency for the model to provide correct diagnoses based on incorrect image interpretations, which could have significant clinical implications.”

Klochko says his study’s findings underscore the need for more specialized and rigorous evaluation methods to assess large language model performance in radiology tasks. “Given the current challenges in accurately interpreting key radiologic images and the tendency for hallucinatory responses, the applicability of GPT-4 Vision in information-critical fields such as radiology is limited in its current state,” he says.

Featured image: Bar graph shows the distribution of answers provided by GPT-4 with vision (GPT-4V; OpenAI), categorized by prompt phrase, as well as the distribution of correct answers.