Abstract 182P
Background
PD-L1 expression is a key predictive factor for anti-PD-1/PD-L1 therapies. This study introduces a deep learning-based automated framework to accurately measure PD-L1 expression in whole slide images (WSIs) of non-small cell lung cancer (NSCLC). The primary goal is to enhance the precision and reliability of the Tumor Proportion Score (TPS)—a critical metric for determining patient eligibility for immunotherapy.
Methods
The proposed framework follows three main steps: (1) locating tumor-containing image patches, (2) segmenting tumor regions, and (3) detecting the cancer cell nuclei within these regions. TPS is then calculated by comparing positively stained cells to the total viable tumor cells. A dataset from Reference Medicine (Phoenix, Arizona), comprising 66 NSCLC tissue samples, was employed. A hybrid human-machine method was used to label large WSIs. Classification models (EfficientNet, Inception, and Vision Transformer)were trained on patches of size 1000 × 1000 pixels. Segmentation performance was tested on multiple UNet and DeepLabV3 variations, while a pre-trained StarDist model replaced traditional watershed techniques for cancer cell nucleus detection. The performances of classification and segmentation tasks were evaluated using a comprehensive set of metrics. We utilized precision, recall, F1-score, and accuracy as the primary evaluation metrics for the classification task. In the segmentation task, the assessment was conducted using several tests, including the Dice Similarity Coefficient (DSC).
Results
PD-L1 expression was categorized into negative (TPS < 1%), low (TPS 1–49%), and high (TPS ≥ 50%). A Vision Transformer–based classifier achieved an F1-score of 97.54%, and a modified DeepLabV3+ model excelled in segmentation with a DSC of 83.47%. The predicted TPS showed a strong correlation with pathologist-assessed TPS (r=0.9635), and the framework’s overall three-tier classification yielded an F1-score of 93.89%.
Conclusions
This deep learning–driven framework for automating TPS evaluation of PD-L1 expression in NSCLC demonstrated promising performance. It holds potential as a predictive tool, offering clinically meaningful results with greater efficiency and lower cost.
Funding
Has not received any funding.
Disclosure
All authors have declared no conflicts of interest.