1249 Improving CD8⁺cell detection in spatial immunology with generalizable H&E deep learning models trained on diverse clinical scanners

<h3>Background</h3> Robust deep learning (DL) models predicting cell-level biomarker expression on hematoxylin and eosin (H&amp;E) images can enable cost-effective high-throughput biomarker evaluation for immune monitoring and immunotherapy stratification.¹ Model generalizability is essential to real-world implementation but complicated by H&amp;E stain variation.^{2 3} Although scanning systems are known to cause H&amp;E stain variation,^{2 3} their impact on H&amp;E-based DL prediction model performance has not been assessed. Using CD8⁺ T cell prediction as the study model, we evaluated the performance of models trained on images of a triple-negative breast cancer (TNBC) cohort acquired by one or more clinical scanners. We show that incorporating images from multiple scanners significantly improved H&amp;E DL model generalizability. <h3>Methods</h3> Tissue sections of three FFPE TNBC samples were subjected to fluorescent multiplex immunohistochemistry (mIHC) staining of DAPI and CD8 and imaged with Axioscan 7 Slide Scanner (table 1). The same sections were then H&amp;E-stained and imaged separately with four different clinical scanners (table 1, figure 1A). CD8 scoring was transferred from mIHC images to H&amp;E images through image registration. 112×112 pixel H&amp;E image patches of 25,564 cells (1:1 CD8⁺:CD8^-) were obtained, totaling 102,256 patches across four scanners (figure 1B), split 80/10/10 for train/validation/test. DenseNet121-based⁴ models were trained on one or more datasets (table 2) with image augmentation and Macenko normalization,⁵ where applicable. Model performance was evaluated using F1 scores at optimal binary classification thresholds. <h3>Results</h3> Although image augmentation increased single-scanner model generalizability (figure 1C), the all-scanner model (F1: ≥0.87) still outperformed the single-scanner models (F1: 0.34–0.76) (figure 1D). Normalizing test datasets to match the training dataset’s color profile did not improve single-scanner model performance (figure 1E). Reduced performance cannot be attributed to differences in pixel resolution as the m-S2 and m-S3 models, whose datasets have highly similar pixel resolution (table 1), did not fare better on each other’s datasets (figure 1D,E). Marked improvement was observed when images from more than one scanner were included in training datasets, especially when the training set included images from the same scanner as the test dataset (figure 1F). <h3>Conclusions</h3> Our study demonstrates that using images from multiple scanners is crucial for the generalizability of H&amp;E DL models, as scanner-associated variations cannot be fully addressed by color normalization and image augmentation. Our findings contribute to current knowledge of developing robust, clinic-ready H&amp;E DL models as cost-effective IHC staining alternatives. <h3>Acknowledgements</h3> This work was supported by the Institute of Molecular and Cell Biology (IMCB), Bioinformatics Institute (BII), Singapore Immunology Network (SIgN), and Agency for Science, Technology and Research (A*STAR). This work was partly funded by MOH-OFYIRG23jan-0021. <h3>References</h3> Shafi S, Parwani AV. Artificial intelligence in diagnostic pathology. <i>Diagn Pathol</i>. 2023;<b>18</b>:109. Tellez D, Litjens G, Bandi P, <i>et al</i>. Quantifying the effects of data augmentation and stain color normalization in convolutional neural networks for computational pathology. <i>Medical Image Analysis</i>. 2019;<b>58</b>:101544. Aubreville M, Stathonikos N, Donovan TA, <i>et al</i>. Domain generalization across tumor types, laboratories, and species — insights from the 2022 edition of the mitosis domain generalization challenge.<i> Medical Image Analysis</i>. 2024;<b>94</b>:103155. Huang G, Liu Z, van der Maaten L, <i>et al</i>. Densely connected convolutional networks. arXiv. 2018;1608.06993. Macenko M, Niethammer M, Marron JS, <i>et al</i>. A method for normalizing histology slides for quantitative analysis. 2009 IEEE international symposium on biomedical imaging: from nano to macro. 2009:1107–1110. <h3>Ethics Approval</h3> The study was approved by the Agency for Science, Technology and Research (A*STAR) Human Biomedical Research Office (A*STAR IRB: 2021-188). <h3>Consent</h3> De-identified patient data was used in our work. Samples were collected with consent from patients.