Fig. 2
Entity key classification performance for chest CT and brain MRI reports. LLM without BERT extracted prompt refers to the modeling process without using BERT’s extraction information as an additional prompt string, and LLM with BERT extracted prompt uses BERT’s extraction information in the input prompt. Accuracy was calculated using the total evaluation set (n = 360) for both chest CT and brain MRI, recall was calculated using reports containing at least 1 lesion for chest CT (n = 311) and 1 diagnosis for brain MRI (n = 310), and specificity was calculated using reports that contained none of the 11 lesions (n = 49) or 5 diagnosis groups (n = 50).
