Objective:To explore whether extracting nuclear features from digital pathology images is more representative than extracting whole-image features and evalnates the effectiveness of both methods in the early diagnosis and prognosis prediction of tumors,providing new strategies for tumor diagnosis and prevention.Methods:A total of 374 breast cancer cases diagnosed between March 2011 and March 2015 at the Affiliated Hospital of Southwest Medical University were collected.Pathological image features were extracted and analyzed in relation to prognostic factors such as tumor pathology type,histological grade,recurrence,metastasis,and survival status.Results:Nuclear features extraction demonstrated better fractal dimension stability and greater robustness to noise than threshold-based segmentation,with significant differences(P<0.05).The best segmentation performance was achieved with an image size of 1 024×1 024 and a magnification of 63×.Nuclear features also demonstrated higher accuracy in tumor WHO staging(AUC=0.550,F1=0.690)and metastasis identification(AUC=0.600,F1=0.890),while both methods performed similarly in tumor recurrence(AUC=0.610,F1=0.930)and survival status prediction(AUC=0.710,F1=0.790).Conclusion:Nuclear feature extraction demonstrates greater representativeness of relevant features and stronger noise resistance compared to traditional threshold segmentation methods.It exhibits higher diagnostic accuracy in breast cancer diagnosis and prognosis prediction,establishing it as a reliable indicator with significant clinical application value.
digital pathology imagesfractal dimensionfeature extractionmachine learning
NETL
NSTL
万方数据
