间质性肺疾病多模态诊断模型的构建和价值OA

Objective To develop a multimodal aggregator-based decision support system for the diagnosis of in-terstitial lung diseases(ILDs)and evaluate its diagnostic performance.Methods Clinical data,chest CT images,and digitized lung pathology slides from 409 patients with ILD were retrospectively collected and categorized into seven classes:usual interstitial pneumonia(UIP),non-specific interstitial pneumonia(NSIP),inhalational-interstitial lung diseases(IN-ILD),other-idiopathic interstitial pneumonias(other-IIP),autoimmune-related interstitial lung diseases(AR-ILD),granulomatous lung disease(GLD),and other-interstitial lung diseases(other-ILD).A deep multimodal learning model was constructed to integrate clinical,imaging and pathological features.For each class,accuracy(ACC),area under curve(AUC),sensitivity,and specificity were calculated.Attention heatmaps were generated to visualize the regions most relevant to the model's predictions.Results Class-wise ACC ranged from 0.800 to 0.983,and AUC ranged from 0.874 to 0.986.UIP,other-ILD,and other-IIP showed the highest diagnostic performance,with most classes achieving an AUC>0.900.Heatmaps indicated that the model mainly focused on fibrotic foci,in-flammatory infiltrates,and granulomatous lesions.Conclusion A multimodal model integrating digitized pathology,imaging,and clinical information can achieve high performance and reasonable interpretability in multi-class ILD clas-sification.It may provide a novel approach for individualized assessment and management of ILD patients and lay a solid foundation for the development of large-scale multimodal diagnostic models for ILD.