ChatGPT and its revolution in the world of radiology

ChatGPT is revolutionizing the world of radiology

Artificial intelligence can now write radiology text almost as well as a radiologist.

Artificial intelligence has demonstrated its potential and is increasingly used in the field of radiology. Generative pre-trained transformer (GPT) models have the potential to completely revolutionize radiology — a revolution comparable only to the milestones of creating the devices that acquire medical images in the first place. This new revolution will improve accuracy, efficiency, decision-making and patient exam outcomes.

ChatGPT and its revolution

ChatGPT is a technology backed by OpenAI, headquartered in San Francisco, United States, and it is a prototype artificial intelligence chatbot specialized in general-context dialogue. The chatbot is a large natural language model, tuned with both supervised and reinforcement learning techniques. It has roughly 175 billion parameters that have to be adjusted during deep-learning-based training, and its computational requirements can only be met in the largest data centers of Azure, Microsoft’s cloud. Since version 3.5 of ChatGPT was released in November 2022, there has been an explosion of uses for this technology — among them, supporting the work of radiology departments.

ChatGPT as linguistic support for radiologists

A recent opinion piece published by the Radiological Society of North America (RSNA) suggests that “the use of natural language processing with tools such as ChatGPT has the potential to revolutionize the field of medical writing” — see the original piece. Radiology reports in particular — fragments of structured or unstructured text, written in natural language, describing the procedure, patient context, findings and conclusions and/or recommendations — are an obvious case for automating medical writing tasks with ChatGPT.

Once the radiologist has identified the diagnosis, they could automate the generation of the report with AI assistance. For example, if the diagnosis is a right frontal lobe meningioma on a brain MRI, the AI could produce the complete report template, such as:

*“*Patient information: Name, age, gender, clinical history.

Examination technique: Contrast-enhanced brain magnetic resonance imaging (MRI) study.

Findings:

– Well-circumscribed, rounded, homogeneous lesion approximately 2 cm in diameter, with heterogeneous enhancement and thick concentric rings, in the right frontal lobe of the cerebral cortex, suggestive of a meningioma.

Discussion:

– Meningioma is a well-known intracranial neoplasm, which suggests a recommendation for early management to prevent growth and possible symptoms from the cerebral mass.

– The heterogeneous enhancement of the lesion is a typical finding in meningiomas, which may mean the differential diagnosis includes other intracranial lesions with a similar enhancement pattern, such as gliomas or inflammatory lesions.

Conclusions:

– Meningioma in the right frontal lobe of the brain, with heterogeneous enhancement and evidence of thick concentric rings; early surgery should be considered to prevent tumor growth and possible neurological symptoms. Consultation with a neurosurgery or radiotherapy specialist is suggested for management.“

Or if the summarized diagnosis is “solitary pulmonary nodule in the left lung associated with moderate pleural effusion,” the report suggested by the AI is:

“Patient information: Name, age, gender, clinical history.

Examination technique: Non-contrast computed tomography (CT) study of the chest.

Findings:

– Presence of a solitary pulmonary nodule in the upper lobe of the left lung, measuring approximately 3 cm in diameter.

– Moderate pleural effusion in the left hemithorax, occupying approximately one third of the pleural space.

– No other abnormalities of the lung parenchyma, hila or great vessels were observed in this study.

Discussion:

– The solitary pulmonary nodule in the left lung is an important finding that raises the possibility of malignancy. Follow-up with imaging studies and possible CT-guided biopsy is recommended to establish a definitive diagnosis.

– The pleural effusion suggests an accumulation of fluid in the pleural cavity, which should be evaluated to determine its underlying cause, which could be secondary to the pulmonary nodule or a separate condition.

– The combination of the pulmonary nodule and the pleural effusion requires further evaluation, including clinical studies and possibly a biopsy of both findings.

Conclusion:

– Solitary pulmonary nodule in the left lung and moderate pleural effusion in the left hemithorax, requiring follow-up and further evaluation to determine the underlying cause and the treatment required. Consultation with a pulmonary specialist and/or an oncologist is recommended for management.“

More advantages than drawbacks

Note that both examples were generated in tests run by the author with version 3.5 of ChatGPT. You may be surprised by how fluent the AI-generated reports are given only the diagnosis as a prompt, but you might also take issue with the precision of the language. Setting aside the merits or flaws of the example above, one thing is clear: the final content of the radiology report is the radiologist’s responsibility, and these suggested templates should be taken for what they are — suggestions. In the end, the radiologist has the authority to edit and improve the report templates, which is without doubt far more efficient than writing the text from scratch, or recording the narration and then paying to have the audio transcribed.

This shows that this is an efficient and reliable use case today, suitable for inclusion as support for radiologists’ narrative work (you can try this functionality, already available in NOVA Enterprise). Even so, while this kind of technology has no clearly defined scope of application, delegating responsibility of a higher ethical order is not feasible at present. In other words, taking over the radiologist’s role in terms of interpretation is unthinkable with current technology.

Yiqiu Shen, a research associate at the New York University Center for Data Science, commented in an opinion editorial published recently in the prestigious journal Radiology that:

“In general, it is fine to use ChatGPT as a linguistic aid or to provide a template, but it is dangerous to rely on ChatGPT to make a clinical decision”

To clarify that point, Shen notes it would be risky to give the AI instructions such as “given the patient’s lab test result, generate a diagnosis and write a report about it”. Instructions of that kind can be followed by AI systems like ChatGPT, but they sit at the boundary of ethical responsibility and should not be considered for the time being.

Thinking about the future, or thinking about the present

Despite its challenges and risks, when used effectively in bounded contexts such as linguistic support for writing radiology reports, ChatGPT’s potential vastly outweighs its current limitations. This technology can significantly reduce the exhausting workload radiologists currently face, given the increase and spread of diagnostic imaging and the shortage of radiologists in most countries. It is exciting to consider the possibilities that ChatGPT and similar AI-based technologies hold for radiology in the present and in the future. We are eager to evaluate the new version 4 released two weeks ago — among many things, it can now interact not only with text but with images. Watch for the next installment…