Document resource
Background Highly multiplexed protein imaging is a critical tool for understanding the tumor microenvironment, yet current cell annotation methods are resource-intensive and lack transferability due to batch effects, relying heavily on manual effort. Thus, developing a machine learning framework for rapid, precise, and generalizable cell-type identification from spatial proteomics data is a critical need.Methods We developed KADANN (Kolmogorov-Arnold Domain Adversarial Neural Network), a framework combining a Kolmogorov-Arnold Network (KAN) and Domain Adversarial Neural Network (DANN), to improve model generalization across different tissue types. 1 Its architecture features an expressive feature extractor built with KAN layers and a domain classifier coupled with a Gradient Reversal Layer to learn domain-invariant features. Our final model is an ensemble combining predictions from the multiclass KADANN with a specialized binary classifier that distinguishes a heterogenous ‘Other’ cell class, using a logit-based blending strategy to refine accuracy. The training was optimized using the Optuna framework to maximize a macro F1-score, employing a hybrid objective function with cost-sensitive focal loss to mitigate class imbalance, and contrastive learning to enforce a more discriminative feature space (figure 1A). To assess performance, we validated KADANN on diverse multiplex immunofluorescence (mIF) datasets across multiple mIF panels and tumor types including lung adenocarcinoma, melanoma, and Merkel cell carcinoma which had been imaged and processed through our AstroPath pipeline.2 Results KADANN demonstrates superior transferability, achieving an average macro F1-score of 0.756 when adapting between disparate datasets ( figure 1B-D). This represents a significant improvement over prior architectures like MAPS or Cellsighter, which reported transferability scores of 0.6.3 The ensemble model consistently outperformed individual models, particularly in delineating ambiguous cell classes, and showed robust generalization across multiple cancer types with minimal domain-specific fine-tuning.Conclusions KADANN establishes a new benchmark for generalizable AI in computational pathology. By integrating KAN layers with a sophisticated adversarial and contrastive learning strategy, our model achieves state-of-the-art performance and transferability. Its architecture provides a direct framework for creating robust tools for digital pathology. For example, we used KADANN to enable the generation of fully-automated tumor-stroma boundary annotations ( figure 2A-C); by classifying all cells in a whole-slide image, KADANN can produce precise tumor borders that are resilient to variations in tissue preparation and scanning. Future work will focus on extending the KADANN framework by incorporating spatial context using graph neural networks (GNNs) and exploring more advanced semi-supervised techniques, such as iterative pseudo-labeling, to further enhance model accuracy and robustness.References Liu Z, Wang Y, Vaidya S, Ruehle F, Halverson J, Soljačić M, Hou TY, Tegmark M. KAN: kolmogorov-arnold networks. arXiv. 2024; arXiv:2404.19756.Berry S, et al. Analysis of multispectral imaging with the AstroPath platform informs efficacy of PD-1 blockade. Science. 2021;372:eaba2609.Shaban M, Bai Y, Qiu H, et al. MAPS: pathologist-level cell type annotation from tissue images through machine learning. Nat Commun. 2024;15:28.Ethics Approval The study was performed in accordance with Johns Hopkins University IRB approval. The study cohort was from Johns Hopkins and used tissue collected under protocols # NA_00085595 and 00090257.Abstract 1110 Figure 1A: Overall KADANN pipeline. B-D: Confusion matrix for Melanoma, Merkel Cell Carcinoma and Lung adenoma samplesAbstract 1110 Figure 2(A) Comparison of tumor border annotations on a multiplex immunofluorescence image. (B) KADANN Generated. (C) Manually annotated