Author: Rico Castillo, Katherine Gonzalez, Casey C. Heirman, Kyle J. Lafata, Casey Y. Lee, Xiang Li, Yvonne M Mowery, Yvonne M Mowery, Daniel Murphy, Allison Pittman, Ashlyn G. Rickard 👨🔬
Affiliation: Duke University, Department of Radiation Oncology, Duke University, University of Pittsburgh 🌍
Purpose: To evaluate the ability of a deep learning model to identify pathomic features in lymph nodes of preclinical head and neck squamous cell carcinoma (HNSCC) models as surrogates for predicting radiation therapy response.
Methods: A deep learning algorithm was tested on mice induced to express two oral cavity tumor models: immunologic human papillomavirus (HPV)-negative MOC1 tumors or poorly immunologic HPV-negative MOC2 tumors, receiving chemoradiation therapy: cisplatin (5 mg/kg, intraperitoneal) and 8 Gy irradiations on days 0 and 7. Mice survival was continuously monitored and later assessed via Kaplan-Meier survival curves. Hematoxylin & Eosin (H&E)-stained whole slide images (WSI) were prepared for LN specimens. For each WSI, ten square ROIs were selected using QuPath software with each ROI covering 1% of the total LN area. The ROIs were processed through the deep learning algorithm to detect lymphocytes. Sixteen topological features characterizing lymphocyte distribution were extracted from the pipeline. Excluding outliers, each feature value was averaged across ROIs and normalized to z-scores. Independent t-tests were conducted on each feature to identify features differentiating MOC1 and MOC2 groups.
Results: Significant differences (p<0.05) were found between MOC1 and MOC2 groups in nine pathomic features, including cluster score, average distance, diameter, average degree, number of nodes, mean K-core, number of K-core, average betweenness, and central node dominance. Kaplan-Meier survival analysis confirmed significantly (p<0.05) higher survival in the MOC1 group compared to MOC2, suggesting the feasibility of utilizing these pathologic features, identified by the deep learning pipeline, on predicting treatment outcomes.
Conclusion: The deep learning pipeline identified nine pathomic features that capture differences in immunological phenotypes of lymph nodes in pre-clinical mice models, demonstrating positive associations with survival rates.