Background:
The response of head and neck squamous cell carcinoma (HNSCC) to radiotherapy varies considerably depending on human papillomavirus (HPV) status, tumor stage, and anatomical location. Although many tumors initially respond to treatment, a substantial fraction are intrinsically radioresistant or acquire resistance over time. In this preclinical study, we examined whether radiomic analysis of nanoparticle contrast-enhanced CT images, termed nano‑radiomics, can distinguish radiation-resistant from radiation-responsive HNSCC tumors.

Hypothesis/Objective:
We developed and evaluated a three-dimensional, sliding-window–based patch-extraction strategy to generate spatial maps that differentiate radiosensitive and radioresistant tumor phenotypes.

Methods:
In vivo studies were performed in genetically engineered mouse models (GEMMs) of HNSCC with spatially and temporally restricted Keap1 loss of function in the oral epithelium. To recapitulate the human disease more faithfully, we have done this in the background of mice with conditional activation of Pik3ca and inactivation of Trp53, two common genomic drivers of human HNSCC [1]. Primary oral tumors were induced in Keap1-proficient (K5TP; radiosensitive) and Keap1-deficient (K5TPK; radioresistant) GEMMs using 4-hydroxytamoxifen and tobacco carcinogen benzo[a]pyrene injections into the oral buccal mucosa. A cohort of eleven HSNCC tumor-bearing mice, with a mean tumor size of 284±135 mm3, including six K5TPK and five K5TP mice, underwent nanoparticle contrast-enhanced computed tomography (nCECT) four days after intravenous administration of a liposomal-iodine nanoparticle CT contrast agent. Tumors were delineated and segmented from nCECT images, and a three-dimensional sliding-window technique was employed to extract patches containing tumor regions. 1400 radiomics features were extracted from each patch to characterize image texture using PyRadiomics [2, 3]. The dataset of patches was divided into training and testing subsets in a 75/25 ratio, maintaining case-level separation to prevent data leakage. The Synthetic Minority Over-Sampling Technique (SMOTE) was utilized to balance the dataset. Feature selection was performed using the Random Forest variable importance method, followed by the training and testing of four classifiers: Support Vector Machine (SVM), K-Nearest Neighbors (KNN), Naïve Bayes, and Linear Discriminant Analysis (LDA). The predictions were integrated with the original CT images to generate visual representations that distinguish radio-sensitive from radio-resistant tumors.

Results:
Extensive optimization resulted in the selection of a 32×32×32 voxel sliding window with a stride of 16 voxels, producing a total of 1,993 patches. Subsequently, 1,400 radiomics features were extracted from each patch. Using the Random Forest feature importance method, the top 20 features were selected. The training area under the curve (AUC) values for the SVM, KNN, Naïve Bayes, and LDA models were 0.81, 0.95, 0.64, and 0.69, respectively. The KNN model, having achieved the highest training AUC, was selected for further analysis. On the testing set, the KNN model yielded an overall AUC of 0.87, an accuracy of 0.77, a sensitivity of 0.74, and a specificity of 0.94. Mean tumor CT intensity was not significantly different between radiosensitive and radioresistant tumor groups (184±71 HU vs. 208±47 HU), thus demonstrating the limitation of the conventional CT metric. Probability maps were overlaid on axial, coronal, and sagittal planes, enabling the differentiation of radiosensitive and radioresistant tumor phenotypes when the KNN model was applied to the test case scenario.

Conclusion:
Radiomic analysis of nanoparticle contrast-enhanced CT images (nano-radiomics) enabled differentiation of radioresistant HNSCC tumors from radiosensitive HNSCC tumors. Use of a sliding window approach enabled intra-tumor spatial interrogation and 3D mapping of radiation sensitivity.

Image/Figure:
Click to view full size
Click to view full size

Image/Figure Caption:
Representative spatial probability maps of two example test cases generated using the proposed 3D sliding window radiomics framework overlaid on a micro-CT image. The top row shows (a) Axial, (b) Coronal, and (c) Sagittal views of a radiosensitive tumor, respectively, while the bottom row shows (d) Axial, (e) Coronal, and (f) Sagittal views of a radioresistant tumor. Color-coded overlays indicate the mean patch-level probability of radiosensitivity (scale: 0–1), projected back onto the tumor region. Warmer colors (red/yellow) indicate higher probability of radiosensitivity, whereas cooler colors (blue) indicate lower probability of radiosensitivity, thus indicating radioresistant regions

Author
Warid Islam
Baylor College of Medicine