Estimation of histopathological types from breast MRI findings using a large language model

Large language models (LLMs) may hold the potential to infer pathological diagnoses from imaging findings such as computed tomography or magnetic resonance imaging (MRI). This retrospective study investigates whether an LLM can accurately predict histopathological types of breast lesions based on descriptive findings in contrast-enhanced breast MRI reports written in natural language.

We retrospectively analyzed findings from diagnostic imaging reports of consecutive cases of contrast-enhanced breast MRI performed between January and December 2024. Textual descriptions of imaging findings were entered into an LLM (OpenAI o3), and its predictions of histopathological types were compared with the actual pathological diagnoses.

A total of 186 lesions from 180 patients were classified into 10 histopathological types, including lesions with combinations of these types. The LLM o3 generated predictions for eight of these types. For the most prevalent type, invasive ductal carcinoma (123 lesions), the model achieved 83.7% sensitivity, 60.8% specificity, a positive predictive value (PPV) of 76.9%, and a negative predictive value (NPV) of 70.6%. For the second-most frequent type, ductal carcinoma in situ (38 lesions), the model achieved 57.9% sensitivity, 89.6% specificity, PPV of 56.4%, and NPV of 90.2%.

The latest LLM accurately inferred invasive ductal carcinoma with typical breast MRI findings, supporting the hypothesis that LLMs can predict histopathological types based on imaging descriptions.