Locally advanced head and neck squamous cell carcinomas (HNSCC) are commonly treated with radiochemotherapy, but improving treatment efficacy and patient quality of life remains a challenge. Reliable preclinical models are essential for testing new therapeutic strategies, biomarkers and understanding radiation response. In this work, we utilized histopathological imaging and unsupervised learning techniques to cluster histopathological slides of 10 HNSCC xenograft models, and assessed how the results correlate with radiosensitivity data from previous in-house experiments.

In prior experiments, athymic nude mice were xenotransplanted with ten HNSCC tumor models of varying radiosensitivity. These models, ordered from radiosensitive to radioresistant based on tumor control dose 50% (TCD50) values obtained from fractionated radiotherapy, include: XF354, UT-SCC45, SAT, UT-SCC14, UT-SCC8, UT-SCC15, CAL-33, FaDu, SAS and UT-SCC5. A total of 71 hematoxylin and eosin (H&E) whole slide images (WSI) were obtained from excised untreated control tumors (4-10 slides per model). For each WSI, we considered the entire tissue area, from a resolution of 0.44 µm x 0.44 µm, to extract tiles of size 224x224. A publicly available deep-learning model pretrained on large histopathological datasets (UNI) was used to extract 1024 deep features from each tile. The features were standardized using Z-score normalization. Dimensionality reduction was then performed using principal component analysis (PCA). The reduced features were subsequently clustered using Gaussian Mixture Models (GMM) into five clusters, a number determined empirically. We calculated the percentage of each cluster within individual slides to compare the clustering distribution across the tumor models, and assessed the correlation with the TCD50 values to examine how they align with radiosensitivity.

Clustering results showed a relatively good visual separation of the clusters in the 2-dimensional space (PCA), in addition to similar cluster distribution within the slides of each tumor model. Visual inspection of the tiles revealed that one cluster likely represents necrotic and stromal areas, while another corresponds to keratin presence in the tissue. The remaining clusters represent tumor regions, though no definitive distinctions were made. Notably, the median percentage of the keratin cluster in each model showed a negative correlation with the TCD50 values, with a Spearman coefficient of -0.75 (p-value=0.01).

Unsupervised clustering shows promise for HNSCC xenograft models profiling, offering a potential complement to the TCD50 information. Future work will investigate feature interpretability, and undergo further validation and clinical relevance assessment.