Abstract
PURPOSE: To apply natural language processing (NLP) to a large volume of structured radiology reports in the investigation of CT imaging features of new liver metastases from primary genitourinary cancers. METHODS: In this retrospective study, a previously reported NLP model was applied to consecutive structured CT reports from 2016 to 2022 to predict those patients with primary genitourinary cancer who developed liver metastasis. Pathology or imaging follow-up served as the reference standard for validating NLP predictions. Subsequently, diagnostic CTs of the identified patients were qualitatively assessed by two radiologists, whereby several imaging features of new liver metastasis were assessed. Proportions of the assessed imaging features were compared between primary genitourinary cancers using the Chi-square or Fisher's exact test. RESULTS: In 112 patients (mean age = 72 years; 83 males), the majority of new liver metastases were hypovascular (73.2%), well defined (76.6%), homogenous (66.9%), and without necrotic/cystic component (73.2%). There was a higher proportion of iso- to hyperdense liver metastases for primary kidney cancer vs other primary genitourinary cancers (42.5% in kidney cancer; 2.3% in ureter/bladder cancer, 8% in prostate cancer, and 0% in testicular cancer; p < 0.05) and a higher proportion of new liver metastases with ill-defined margin for primary prostate cancer vs other primary genitourinary cancers (44.0% in prostate cancer, 15.0% in kidney cancer, 18.6% in ureter/bladder cancer, and 25.0% in testicular cancer; p < 0.05). CONCLUSION: New liver metastases from primary genitourinary cancers tend to be hypovascular and show several distinct imaging features between different primary genitourinary cancers.