卵巢癌转移灶的智能识别与结构化报告填充:一项多中心研究OA

Objective To explore application models of artificial intelligence for precise localization and evaluation of ovarian cancer metastases.Methods A total of 273 contrast-enhanced abdominal-pelvic com-puted tomography(CT)scans from patients with ovarian cancer metastases across three centers were includ-ed.Radiologists annotated 174 subdiaphragmatic metastases and 516 perihepatic metastases,who were ran-domly divided into training(n=561)and test(n=129)sets.A deep convolutional network-based binary classification model for distinguishing subdiaphragmatic/perihepatic locations was constructed,and its accura-cy,sensitivity,specificity,precision,F1-score,and area under the curve(AUC)were calculated.Using four-phase contrast-enhanced CT images and surgical pathology data,the structured reports were filled in and their performanc eevaluated.Results The subphrenic/perihepatic location differentiation model achieved an AUC of 0.78,with an accuracy of 0.721,sensitivity of 0.417,specificity of 0.839,precision of 0.500,and an F1-score of 0.455.In the structured report population,the classification model for perihepatic metastasis location performed best,attaining an AUC of 0.83,accuracy of 0.753,sensitivity of 0.804,specificity of 0.702,preci-sion of 0.725,and an F1-score of 0.763.The recognition capabilities of models for other features require fur-ther improvement.Conclusion This work establishes a novel clinical assistance workflow—"automated image analysis,key feature extraction,and structured report population"—offering a practical framework for optimiz-ing diagnostic processes and enhancing reporting standardization.