{
  "abstract": "Background Tumor-infiltrating lymphocytes (TILs) are key indicators of anti-tumor immunity and patient prognosis across multiple cancers. Traditional assessments using H&E slides rely on subjective visual estimation and often ignore spatial context. Recent studies distinguish lymphocytes in tumor and stromal regions, termed intratumoral TILs (iTILs) and stromal TILs (sTILs), respectively. Such spatial distribution and compartmental localization may offer complementary prognostic value. We present a deep learning framework for estimating densities of iTILs, sTILs, and tumor cells at the patch level to develop a prognostic stratification scheme of tumor-immune interactions from a TIL perspective.Methods We used 10x Genomics public Xenium datasets across 9 organs (24 samples) with aligned H&E images, totaling 202,605 image patches (256×256 at 40× magnification). H&E images were registered to Xenium coordinates using Valis and normalized with Reinhard color normalization. Cell types were annotated via anchor-based methods. Tumor boundaries were computed using Delaunay triangulation ( figure 1). Immune cells inside tumor boundaries or within a 100 μm margin were defined as iTILs, while those beyond were classified as sTILs. For each patch, we computed log-transformed densities of iTILs, sTILs, and tumor cells. We trained both single-task and multitask models to predict these densities using deep learning architectures including EfficientNetV2,1 ConvNeXt,2 ResNet-50,3 Swin Transformer,4 and CONCH5 with regression heads.Results In single-task models, the best iTIL prediction was achieved with EfficientNetV2 (Spearman = 0.8127; R 2 = 0.7532), while ConvNeXt performed best for sTIL (Spearman = 0.8006; R2 = 0.7743). EfficientNetV2 also performed well for tumor cell density (Spearman = 0.8041; R2 = 0.7620). Multitask training is ongoing.To assess clinical utility of TIL spatial patterns, we explored density, number, size, and distance of TIL clusters. In 6 responders and 3 non-responders to immune checkpoint blockade therapy for colorectal cancer, responders showed higher mean TIL density (+0.26), more clusters (+2.17), larger sizes (+5.09), and shorter inter-cluster distances (−0.95), suggesting spatial heterogeneity of immune activity as a potential prognostic marker.Conclusions Our study shows that spatially resolved, patch-level prediction of iTIL, sTIL, and tumor cell density can serve as a prognostic tool in immuno-oncology. Unlike binary classification of TILs, our Methods leverages continuous, spatially contextualized signals enabled by Xenium data. The iTIL vs. sTIL distinction, rarely explored previously, may yield insights into tumor-immune interactions. We anticipate that multitask learning may improve generalizability. Application to larger clinical datasets is planned for validation and translational use.References Tan M, Le Q. EfficientNetV2: smaller models and faster training. Proc Mach Learn Res. 2021;139:10096–10106.Liu Z, Mao H, Wu C-Y, Feichtenhofer C, Darrell T, Xie S. A ConvNet for the 2020s. Proc IEEE/CVF Conf Comput Vis Pattern Recognit. 2022.He K, Zhang X, Ren S, Sun J. Deep residual learning for image recognition. Proc IEEE Conf Comput Vis Pattern Recognit. 2016.Liu Z, Lin Y, Cao Y, et al. Swin transformer: hierarchical vision transformer using shifted windows. Proc Int Conf Comput Vis. 2021.Mahmoodlab. Towards a visual-language foundation model for computational pathology (CONCH). arXiv. 2023;2307:12914.Abstract 1097 Figure 1Preprocessing of Xenium and H&E data. TIL density prediction regression model structure",
  "authors": [
    {
      "affiliations": [
        "Ewha Womans University, Seoul, Republic of Korea"
      ],
      "name": "Minkyung Kim"
    },
    {
      "affiliations": [
        "Ewha Womans University, Seoul, Republic of Korea"
      ],
      "name": "Sanghyuk Lee"
    }
  ],
  "title": "1097 AI-powered spatial quantification of intratumoral TILs for pan-cancer prognostic stratification",
  "uid": "16cf56cf-47c0-58e9-ba1a-79a81cb9179e"
}