A feature extraction method and system based on deep learning and radiomics

Abstract

The application relates to the technical field of image intelligent analysis, and specifically discloses a feature extraction method and system based on deep learning and imageomics, which comprises the following steps: through calculating the overlap rate of deep learning and imageomics ROI boundaries and the absolute difference of Z-score standardization of GLCM-entropy, double determination and accurate identification of ROI boundary double inconsistency are realized; based on a spatial attention mechanism, a pixel weight distribution model is constructed in combination with texture abnormality, abnormal weight elimination, smoothing normalization processing and adaptive filtering of pseudo boundaries are carried out, a unified coordinate system with the imageomics boundary center as the origin is re-constructed, image standardization is realized by minimizing and optimizing affine transformation parameters through the Hausdorff distance, finally, a three-order Bezier curve is used to fit the dislocation area and iterative smoothing correction is carried out, high-precision unified ROI boundaries are established, image correction and efficient fusion of heterogeneous features are realized, and the effectiveness of coupled feature representation is improved.

Images