The invention discloses a decayed tooth semantic segmentation method based on structure
perception, and belongs to the field of oral clinical
image analysis. The invention provides a segmentation method combining anatomical structure constraint and dynamic training regulation for solving the problem of supervision bias caused by
low contrast of a dental caries focus, remarkable background structure interference and a minimum target. The method comprises the following steps: firstly, constructing a continuous space proximity field taking a focus as a center, and quantifying anatomical proximity of a background and the focus; and on the basis of a multi-scale fuzzy hierarchy generation mechanism, a progressive fuzzy background of a continuous anatomical gradient is generated through fuzzy layer interval selection and pixel-level weighted fusion so as to simulate a visual attenuation rule of dental caries focus centralization and weaken high-similarity gray
background noise. In the training stage, a four-stage focusing gradient weight scheduling strategy is constructed, so that the model is gradually transited from global anatomical
structure learning to
focus area strengthening. The method does not need to change the basic structure of an existing segmentation network, and can improve the
dissection consistency and boundary stability of dental caries segmentation.