The invention belongs to the field of myopia bottom-of-eye
lesion prediction, and discloses a myopia bottom-of-eye
lesion progress prediction method and
system based on multi-
modal sequence data, and the method comprises the steps: extracting
fundus image features through employing a
local structure guided multi-scale vision Transform; the method comprises the following steps: acquiring table data, extracting features through respective network modules, and carrying out cross-attention mechanism fusion;
processing data of irregular time points by using Transform time coding and an LSTM pseudo sequence method, and converting the data into continuous
time sequence input; estimating the contribution of each time
point data in prediction by using time interval coding and
time difference weighting; a weighted attention mechanism is adopted to endow
modal information of different time points with different weights, and data features which have the most influence on the progress in different periods are identified; the image features and the table data features are fused through a cross-attention mechanism, and a unified
feature vector is formed; and outputting the progress trend of future lesions of the patient. According to the method, the
lesion occurrence or progress trend at a specific age or a future time point can be predicted, and a scientific basis is provided for individualized management.