Abstract
This work establishes a non-invasive paradigm for mapping immunosuppressive microenvironments, bridging precision radiotherapy with immunotherapy personalization. The Immune Evasion Score (IES) provides a dynamic biomarker of immune evasion, potentially guiding patient stratification for checkpoint inhibitors.
Introduction
Immunotherapy has fundamentally altered the treatment landscape for lung cancer, yet its efficacy remains limited to approximately 20–30% of patients. Current clinical practice relies heavily on PD-L1 immunohistochemistry (IHC) derived from invasive needle biopsies to predict response. However, biopsies are prone to sampling bias due to tumor heterogeneity and provide only a static snapshot of a highly dynamic immune-tumor interaction. There is a critical need for non-invasive, longitudinal biomarkers that can capture the "crosstalk" between tumor growth kinetics and the immune microenvironment.
Methodology: The MK-UNet Framework
To address these challenges, researchers developed MK-UNet, a medical knowledge-guided 3D deep learning framework designed to map the immune landscape of lung tumors using longitudinal CT scans.
Longitudinal Radiomics and Growth Kinetics
Unlike traditional models that analyze static images, MK-UNet integrates the Log Volume Change Rate (LVCR). By analyzing sequential CT scans, the model quantifies tumor growth kinetics, recognizing that rapid proliferation is often a hallmark of immune evasion.
Immune-Aware Attention Gates (IAAG)
The architecture incorporates novel Immune-Aware Attention Gates (IAAG). While standard attention mechanisms focus on salient pixels, IAAG weights features based on both imaging patterns and the LVCR. This ensures the model prioritizes regions characterized by aggressive growth and irregular margins—areas most likely to harbor immunosuppressive niches.
Multi-modal Fusion and the Immune Evasion Score (IES)
At the network's bottleneck, a 6D clinical vector (including age, gender, and smoking history) is injected to provide essential patient context. The final output is the Immune Evasion Score (IES), a composite metric derived from growth rate, tumor sphericity, and predicted PD-L1 distribution, quantifying the tumor's ability to circumvent immune detection.
Results and Validation
The model was trained on an internal dataset of 420 lung CT scans and validated against genomic and imaging data from the TCGA-LUAD and TCIA repositories.
Segmentation and Interpretability
MK-UNet achieved a Dice coefficient of 0.7728, significantly outperforming standard 3D U-Net architectures. Visual analysis of the segmentation process demonstrates the model's ability to refine tumor boundaries through windowing and filtering.
Clinical Predictive Power
The IES demonstrated high accuracy in predicting PD-L1 expression levels, achieving an AUC of 0.85 ($p < 0.001$). Furthermore, high IES scores were associated with a 2.3x increased risk of rapid disease progression during immunotherapy. Notably, the model maintained high efficiency, completing inference in approximately 15 seconds per scan.
Discussion
The integration of tumor kinetics into deep learning architectures represents a significant step toward "virtual biopsies." By demonstrating a 72.3% concordance between AI-prioritized regions and pathological PD-L1+ niches, this study validates the biological plausibility of radiomic-based immune profiling.
While the requirement for longitudinal data (at least two scans) presents a hurdle for immediate treatment scenarios, the MK-UNet framework provides a robust tool for monitoring treatment response and stratifying patients. Future research will focus on multi-center validation and the development of texture-based proliferation scores to enable similar insights from single-timepoint imaging.
Summary updated