STROKE LESION SEGMENTATION METHOD AND SYSTEM

A method and system that improves the accuracy of segmenting a stroke lesion is disclosed. The method establishes a correlation between slices by changing a traditional U-Net model into a long short-term memory (LSTM) U-Net segmentation network model by replacing some convolutional units in the traditional U-Net model with convolutional LSTM units. In a data preprocessing stage, the present disclosure performs ordered rotating slicing operations on three-dimensional (3D) image data of a stroke to generate first and second two-dimensional (2D) forward and reverse ordered rotating slice sequences, and trains a segmentation model by the first ordered rotating slice sequence to generate forward and reverse LSTM U-Net segmentation network models. The method segments the second ordered rotating slice sequence by the forward and reverse LSTM U-Net segmentation models to obtain forward and reverse lesion segmentation result sequences. The method also fuses the data. [FIG. 2] Normal line Sagittal plane (t-1) LSTM U-Net (t-1) (t) LSTM U-Net (t) (t+n) LSTM U-Net (t+1)