This invention discloses a method and
system for dynamic segmentation of
cervical lesion edges, belonging to the field of medical
image processing technology. Local
grayscale truncation and reconstruction, and cross-spectral
reflectivity consistency correction are performed on the registered multispectral panoramic image of the
cervix to obtain a de-reflected multispectral base image and a
reflectivity-reliable
mask. Under the constraint of the
reflectivity-reliable
mask, local fine-scale
pathological texture features and large-scale vascular morphological features are extracted. Orthogonal projection and gated fusion are performed on the large-scale vascular morphological features to obtain a high-dimensional fusion
feature matrix. A stable probability heatmap is obtained through scale
recovery, regional overlap loss, and boundary continuity loss training.
Radial gradient extreme value boundary points are extracted, and constrained contour evolution is performed to generate a continuously closed set of
lesion boundary coordinates. This invention enables the reflectivity-reliable
mask to be used throughout the low-confidence positional direction constraint reconstruction, orthogonal gated fusion, and
lesion boundary generation processes, reducing the
impact of reflectivity
occlusion, pixel imbalance, and boundary breaks on the boundary coordinate output.