Abstract / Summary
Tissue-based assays provide broad screening for neural autoantibodies but remain vulnerable to technical and reader-dependent variability. We evaluated an integrated workflow combining automated 3,3'-diaminobenzidine tissue-based assay processing, digital whole-slide imaging, and interpretable deep learning. Model development used 2,357 retrospective serum and cerebrospinal fluid images referred from more than 80 institutions and processed in a centralized laboratory, with reference labels established by expert-consensus tissue-based assay interpretation. A prospective validation cohort included 200 images from 200 unique patients (154 positive and 46 negative; 100 serum and 100 cerebrospinal fluid). The same images were evaluated by two expert and two novice readers in a fixed-order two-round study, with unaided interpretation followed by artificial-intelligence-assisted interpretation after a washout of at least two weeks. The standalone binary classifier achieved an area under the receiver operating characteristic curve of 0.922, accuracy of 86.5%, sensitivity of 85.7%, and specificity of 89.1% in prospective validation. Pooled reader accuracy was 90.4% in the unaided round and 97.4% in the artificial-intelligence-assisted round (McNemar P < 0.001), with specificity of 82.6% and 99.5%, respectively. Novice-reader accuracy was 83.3% unaided and 96.5% with artificial-intelligence assistance (P < 0.001), approaching unaided expert accuracy of 97.5%. Among tissue-based-assay-positive cases, reader accuracy for cytoplasmic/nuclear antigen-associated versus neural surface antigen-associated staining-pattern subclassification was 88.8% unaided and 94.0% with artificial-intelligence assistance (P < 0.001). These findings support further multicentre, cross-platform evaluation of artificial-intelligence-assisted neural autoantibody tissue interpretation.