ABSTRACT To develop an advanced multimodal deep learning network for prognosis prediction in anti‐N‐methyl‐D‐aspartate receptor (anti‐NMDAR) encephalitis, addressing key challenges such as semantic ambiguity and decreased efficiency in integrating cross‐modal features, including MRI and clinical features. A novel anti‐NMDAR encephalitis prognostic model has been proposed, called PCA‐Net. First, a lesion enhancement sequence based on traditional MRI mode was constructed to more efficient cross‐modal feature interaction. Subsequently, a multimodal collaborative attention module was proposed to promote in‐depth interaction between cross‐modal features. The core of PCA‐Net is a new self‐supervised learning strategy, which is different from existing methods to constrain each single modal feature separately. The proposed method allows the network to explore the relationship between cross‐modal features first, avoiding premature constraints leading to over‐normalization. Evaluations on two independent datasets show that PCA‐Net consistently outperforms existing state‐of‐the‐art methods in the prognosis of patients with anti‐NMDAR encephalitis. By enhancing cross‐modal feature integration, the proposed method contributes to the development of more accurate prognostic tools for anti‐NMDAR encephalitis, with potential implications for improving clinical decision‐making.
He et al. (Sun,) studied this question.