Supervised Learning Assisted by Text-Guided Attention for Ards Prediction

Acute respiratory distress syndrome (ARDS) poses significant challenges in clinical diagnosis due to its complex imaging characteristics and the presence of extraneous noise in CT images. Existing deep learning approaches often struggle with noise interference, necessitating time-consuming manual segmentation processes. To address these limitations, we propose a novel method called text-guided attention (TGA), which leverages CT reports as supervision to guide the focus of the CT image encoder on diagnostically relevant regions, thereby minimizing attention to noisy areas. Additionally, we incorporate an ARDS-related knowledge graph as prior knowledge, enhancing the model's predictive capabilities for both ARDS diagnosis and 28-day survival period estimation through a multi-task learning framework. Experimental results demonstrate that TGA outperforms other methods by a significant margin, highlighting its effectiveness under the non-segmentation condition, with an ACC (diagnosis) improvement of 4.9% and a C-index (survival period estimation) increase of 34.7% compared to the best alternatives.