The invention provides a multi-agent collaborative semantic transformation framework for a
rare disease recognition problem, and belongs to the technical field of
artificial intelligence and medical
image analysis. According to the method, the problems of scarcity of marked data and difficulty in cross-
domain knowledge migration in
rare disease recognition are solved. The method comprises the steps that S1, a multi-agent parallel
network architecture is constructed, each agent is provided with a special semantic focusing module, and diversified features are extracted from different attribute perspectives; s2, a cooperative gating mechanism with dynamic temperature parameters is realized, and the cooperative and competitive relationship between intelligent agents is balanced in a self-adaptive manner; s3, applying a double-constrained cross-domain
semantic alignment strategy to ensure that the converted semantic features are consistent with the original features and keep diversity at the same time; s4, adopting a progressive training strategy and a
semantic consistency loss function to reduce an
overfitting phenomenon in cross-
domain knowledge migration; and S5, classifying and identifying rare diseases through multi-agent cooperation, and applying source
domain knowledge to a target domain. The method is prominent in medical image application, the accuracy rate of identifying rare
skin diseases by using only common
skin disease data reaches 52.13%, and the method is remarkably superior to an existing method. The framework has wide applicability in cross-domain zero sample learning tasks, is particularly suitable for application in the field with definite definition of semantic attributes, and provides a new solution for diagnosis of rare diseases in medical images.