TL;DR
This paper introduces a novel deep learning-based "Pathomics" workflow to create detailed spatial maps of tumor-infiltrating lymphocytes (TILs) within colon cancer tissue. These "Tumor-TIL maps" allow for the quantitative assessment of TIL abundance and distribution, revealing that higher TIL levels are significantly associated with improved patient survival, positioning TILs as a promising prognostic biomarker for precision oncology.
What Problem Does This Paper Address?
Colon cancer remains a significant global health challenge, and while immunotherapy shows promise, identifying patients who will benefit remains difficult. Tumor-infiltrating lymphocytes (TILs) are immune cells that play a crucial role in fighting cancer, and their presence is known to be linked to better outcomes. However, routinely assessing TILs in clinical practice is hindered by time constraints, the need for specialized training, and observer variability. This paper aims to bridge this gap by developing an automated, quantitative method to analyze TILs in standard H&E-stained colon cancer tissue slides, making this valuable prognostic information more accessible for clinical and research applications.
Methods
The core of this work is a deep learning "Pathomics" workflow that leverages two distinct convolutional neural network (CNN) models. The first, a ResNet34-based model, is trained to detect and segment colon cancer regions within whole slide images (WSIs). The second, a third-generation InceptionV4-based model, is used to identify and map lymphocytes (TILs). By combining the outputs of these models and applying a probability threshold, the workflow generates "Tumor-TIL maps." These maps visually represent the spatial distribution of both tumor and TILs, allowing for the calculation of key metrics like "Tumor Content" (tumor percentage) and "Intratumoral TILs%" (the proportion of tumor tissue infiltrated by TILs). The study also employs standard statistical methods, including Kaplan-Meier survival analyses and multivariate Cox regression, to assess the prognostic significance of these computed metrics.
Datasets
The study utilized two primary datasets:
- TCGA COAD dataset: This dataset comprised 458 patients with 411 diagnostic H&E WSIs. After exclusions for image quality and small biopsy sizes, 347 WSIs were analyzed.
- UKentucky dataset: This dataset included 100 patients, with 245 H&E WSIs. After excluding non-cancer images, 178 cancer-containing images were analyzed, with an average TILs% calculated per patient for 58 cases.
Additionally, a private collection of 200 colon cancer H&E WSIs from the University of Pennsylvania was used to generate training data for the colon cancer segmentation model, but these were not included in the final analysis due to a lack of corresponding clinical data.
Key Insights and Ideas
- Visualizing the Immune Landscape: Tumor-TIL maps provide an intuitive and visually rich way to understand the complex spatial interactions between tumor cells and immune cells within the tumor microenvironment.
- TILs as a Quantitative Biomarker: The workflow successfully quantifies TIL abundance and distribution, moving beyond qualitative assessments to provide a robust, data-driven metric (TILs%).
- Prognostic Power of TILs: The study strongly reinforces the idea that higher TIL infiltration is a significant indicator of better patient outcomes, not just in localized but also in metastatic colon cancer.
- Context Matters: The analysis highlights that TILs' prognostic value is influenced by other factors like tumor content and stage, suggesting a nuanced interplay that can be explored with these detailed maps.
Key Results
- High intratumoral TILs% was significantly associated with improved overall survival (OS) and progression-free interval (PFI) in the TCGA COAD cohort.
- This association held true across various clinicopathologic variables, including AJCC stages (both localized I-II and metastatic III-IV), histological grade, and microsatellite instability (MSS/MSI-L).
- In the UKentucky dataset, high TILs% also showed a positive relationship with OS, particularly in localized stages and with smaller tumor sizes.
- Multivariate Cox regression analysis in the TCGA COAD cohort identified high TILs% as a significant predictor of improved PFI (p=0.016), while tumor percentage was a significant factor for both OS (p=0.004) and PFI (p=0.05).
Limitations and Future Work
The authors acknowledge several limitations:
- Limited Sample Size: While the study used two datasets, some subgroup analyses, particularly in the UKentucky cohort and for specific MSI-H cases, were limited by small sample sizes, potentially underpowering some statistical findings.
- Tumor Heterogeneity: The study primarily focused on WSIs from a single diagnostic slide per case, which might not fully capture the tumor heterogeneity present across multiple slides.
- Underrepresentation of TME Complexity: The current workflow focuses on cancer cells and TILs, potentially underrepresenting other crucial components of the tumor microenvironment like stromal reactions, other immune cell types, and microvasculature.
Future work suggested includes:
- Reproducibility and Validation: Further validation with independent and larger cohorts is crucial to confirm the findings and refine the methodology for clinical translation.
- Integration with Other Data Modalities: Combining Pathomics data with genomic, transcriptomic, metabolomic, and pharmacologic data from clinical trials is envisioned to provide a more comprehensive understanding of tumor-immune interactions.
- Advanced Segmentation: Exploring more advanced tools for nuclear segmentation and classification could further enhance the characterization of tumor and immune cells.
- Clinical Implementation: The ultimate goal is to refine this workflow for routine clinical use, enabling personalized treatment planning and guiding immunotherapy selection.
Summary updated