Multi-incentive fusion zero sample lesion detection method based on visual language model

The invention provides a multi-excitation fusion zero sample lesion detection algorithm based on a visual language model, and belongs to the technical field of multi-modal medical image processing. The method provided by the invention comprises the following steps of: 1) directly inputting a plurality of excitations into a model to obtain a corresponding intermediate variable C; 2) selecting a proper fusion strategy, and classifying the intermediate variable C; and 3) performing position clustering, size clustering, category label correction and confidence coefficient threshold screening on the classified intermediate variable C '. And 4) performing multi-stage feature fusion screening on the screened # imgabs0 # from different excitations, and sending the screened # imgabs0 # into a small classification network for further classification judgment to obtain a final fusion result. Through the idea of integrated learning and the assistance of a deep learning basic network framework, the limitation of an origina