Radiologists' little helper in the Open AI Store

It has been almost 3 months now since Open AI launched their GPT store where you can get customized GPTs that are tailored to a specific task. If you are on the premium subscription this feature is available for you.

Intrigued by what's out there for radiology, I decided to skim through the offerings and put a couple to the test - partly to see their real-world utility and partly, admittedly, to prove just how indispensable we radiologists are 😋.

I zeroed in on the top two hits for "Radiology" in the GPT store:

1. Radiologist & Radiology Assistant by unaded.com: https://chat.openai.com/g/g-8m153aROb-radiologist-radiology-assistant
2. Radiology Copilot by Lakshminarayan Srinivasan: https://chat.openai.com/g/g-o6cGAEN3H-radiology-copilot

The First Challenge

The first prompt to assess these GPTs: “My referring orthopedic surgeon has referred a patient for magnetic resonance imaging (MRI) evaluation. The clinical suspicion is an anterior cruciate ligament (ACL) rupture. The surgeon is seeking a comprehensive assessment of any additional intra-articular pathology within the knee joint. What is the best protocol for this task?”

The results:

1. Radiologist & Radiology Assistant:

This one served up details on the MRI sequences to use, explaining their relevance. It even highlighted specific scan areas of interest - pretty handy for beginners. Overall, a decent job.

2. Radiology Copilot:

This GPT takes a different approach: It first recommends the particular planes to use, and then within those planes, it provides more detailed advice on the sequence types I should apply. This method is a bit unusual, but still acceptable. I appreciated its additional protocol suggestions for a more accurate diagnosis and its nod to patient safety and comfort.

The Second Challenge

Next, I threw in a CT image of a lung carcinoma in the left upper lobe.

The task? “What pathology can you detect on the image? How could I explain it with 1 sentence to a) the patient, 2) the referring GP.”

1. Radiologist & Radiology Assistant:

It mixed up the sides but correctly located the upper lobe. The details were somewhat off (e.g., “clear surrounding lung parenchyma”) and incomplete (missing mention of pleural contact).

2. Radiology Copilot:

It identified the correct side but it did not give any further details that could add value to a radiology report.

What I think about it?

The answers to the first prompt were quite useful. Both GPTs offered useful insights on scanning protocols. The Radiology Copilot’s approach was a bit uncommon and sure, their suggestions weren’t a carbon copy of MSK society recommendations, but that’s pretty reflective of clinical practice diversity, right 😊? In the end, their protocols could satisfactorily answer the orthopedic surgeon’s questions.

The second test, though, spotlighted the LLMs' limitations. There were mix-ups and omissions in details. But in explaining to the patient and GP, they both shone, offering near-perfect responses with only minor tweaks needed (like using 'mass' instead of 'solitary pulmonary nodule').