1216 Development of a lung specific foundation model, PERCIVAL- LUNG: predictive enhanced research for cancer investigations via AI learning for LUNG

<h3>Background</h3> Foundation models in digital pathology have proven to improve the efficiency of tasks, such as tumor, tumor subtype, and tumor microenvironment classification and nuclei segmentation, with the model encoder serving as a starting point for these task-specific decoders.^1–3 These models are trained on a wide cohort of images incorporating different cancers, multiple stains and scanners. Despite significant advancements, existing foundation models in digital pathology have not yet fully captured the intricate variability to a particular disease or condition as it involves the careful curation of a dataset capturing the variability across cancers, their subtypes, and morphological heterogeneity. In this study, we introduce our lung foundation model PERCIVAL- LUNG: Predictive Enhanced Research for Cancer Investigations via AI learning for LUNG. With PERCIVAL-LUNG, we narrow the scope of the development of foundation models to applications to problems pertaining to Non-Small Cell Lung Cancer (NSCLC) since it’s the most prevalent form of lung cancer comprising of approximately 85% of all lung cancers.⁴ In our results, we have found that the model learns a robust set of features that captures heterogeneity at the sub-type level and a variety of morphological patterns specific to NSCLC. <h3>Methods</h3> Our model leverages the self-supervised learning strategy presented by DINO⁵ which uses the concept knowledge distillation⁶ on Vision Transformers.⁷ Our approach differs from that of pan-cancer foundation models because we carefully curate a development dataset that is specific to NSCLC ensuring that it captures the variability associated with its subtypes and morphological patterns. Our cohort size has thousands of whole slide images translating into millions of patches that are fed to our foundation model. Each sample in our dataset has metadata associated with NSCLC sub-types and histologic patterns, which underwent quality control (QC) checks from a board-certified pathologist. We leverage this granularity in our model development to ensure the model learns important NSCLC-specific features. <h3>Results</h3> We evaluated our performance on both our internal curated test set and benchmarked it upon TCGA⁸ for tumor and its microenvironment classification, tumor sub-type classification and biomarker classification and observed our model outperformed on these studies when compared to our previous self-supervised contrastive learning-based encoders and without. <h3>Conclusions</h3> Through our foundation model, we aim to enhance the diagnostic and prognostic capabilities in NSCLC pathology and, ultimately, contributing to more personalized and effective patient care. <h3>References</h3> Chen RJ, Ding T, Lu MY, Williamson DF, Jaume G, Song AH, Chen B, Zhang A, Shao D, Shaban M, Williams M. Towards a general-purpose foundation model for computational pathology. <i>Nat Med</i>. 2024;<b>30</b>(3):850–862. Shaikovski G, Casson A, Severson K, Zimmermann E, Wang YK, Kunz JD, Retamero JA, Oakley G, Klimstra D, Kanan C, Hanna M. PRISM: a multi-modal generative foundation model for slide-level histopathology. <i>arXiv</i>. 2024;arXiv:2405.10254. Juyal D, Padigela H, Shah C, Shenker D, Harguindeguy N, Liu Y, Martin B, Zhang Y, Nercessian M, Markey M, Finberg I. PLUTO: pathology-universal transformer.<i> arXiv</i>. 2024;arXiv:2405.07905. Prevalence of non-small cell lung cancer. Medical News Today. Available from: https://www.medicalnewstoday.com/articles/prevalence-of-non-small-cell-lung-cancer Oquab M, Darcet T, Moutakanni T, Vo H, Szafraniec M, Khalidov V, Fernandez P, Haziza D, Massa F, El-Nouby A, Assran M. Dinov2: learning robust visual features without supervision. <i>arXiv</i>. 2023;arXiv:2304.07193. Hinton G, Vinyals O, Dean J. Distilling knowledge in a neural network. <i>arXiv</i>. 2015;arXiv:1503.02531. Dosovitskiy A, Beyer L, Kolesnikov A, Weissenborn D, Zhai X, Unterthiner T, Dehghani M, Minderer M, Heigold G, Gelly S, Uszkoreit J. An image is worth 16x16 words: transformers for image recognition at scale. <i>arXiv.</i> 2020;arXiv:2010.11929. The Cancer Genome Atlas Program (TCGA). National Cancer Institute (NCI).