Text-Guided Cross-Position Attention for Segmentation: Case of Medical Image

We propose a novel text-guided cross-position attention module which aims at applying a multi-modality of text and image to position attention in medical image segmentation. To match the dimension of the text feature to that of the image feature map, we multiply learnable parameters by text features and combine the multi-modal semantics via cross-attention. It allows a model to learn the dependency between various characteristics of text and image. Our proposed model demonstrates superior performance compared to other medical models using image-only data or image-text data. Furthermore, we utilize our module as a region of interest (RoI) generator to classify the inflammation of the sacroiliac joints. The RoIs obtained from the model contribute to improve the performance of classification models.