TGFN-SD: A text-guided multimodal fusion network for swine disease diagnosis

Abstract

China is the world's largest producer of pigs, but traditional manual prevention, treatment, and diagnosis methods cannot satisfy the demands of the current intensive production environment. Existing computer-aided diagnosis (CAD) systems for pigs are dominated by expert systems, which cannot be widely applied because the collection and maintenance of knowledge is difficult, and most of them ignore the effect of multimodal information. A swine disease diagnosis model was proposed in this study, the Text-Guided Fusion Network-Swine Diagnosis (TGFN-SD) model, which integrated text case reports and disease images. The model integrated the differences and complementary information in the multimodal representation of diseases through the text-guided transformer module such that text case reports could carry the semantic information of disease images for disease identification. Moreover, it alleviated the phenotypic overlap problem caused by similar diseases in combination with supervised learning and self-supervised learning. Experimental results revealed that TGFN-SD achieved satisfactory performance on a constructed swine disease image and text dataset (SDT6K) that covered six disease classification datasets with accuracy and F1-score of 94.48 % and 94.4 % respectively. The accuracies and F1-scores increased by 8.35 % and 7.24 % compared with those under the unimodal situation and by 2.02 % and 1.63 % compared with those of the optimal baseline model under the multimodal fusion. Additionally, interpretability analysis revealed that the model focus area was consistent with the habits and rules of the veterinary clinical diagnosis of pigs, indicating the effectiveness of the proposed model and providing new ideas and perspectives for the study of swine disease CAD.