Radiomics-based prediction of radiosensitivity from preclinical HNSCC histopathology images

Purpose/Objective:

Biomedical images are a source of high-dimensional information that can be extracted using feature-based machine learning (radiomics). Moreover, preclinical radiotherapy research using animal models generates an enormous number of images that are generally overlooked as a source of quantitative image-based biomarkers. Here we apply radiomics feature extraction methods to 2D histopathology data from a published animal trial (Rassamegevanon et al., Radiother. Oncol., 2019) to classify HNSCC tumors based on their known radiosensitivity.
 
Materials/Methods:
Athymic mice were xenotransplanted with three HNSCC tumor models of varying radiosensitivity. The tumors were irradiated in vivo with 2 – 8 Gy, excised 24 hours post-treatment, and stained for nuclear marker DAPI. Up to 13 ROIs per tumor were imaged using fluorescence microscopy and the nuclei were segmented using ImageJ plugin StarDist. Standardized radiomic features were extracted and clustered using the radiomics processor MIRP. Then, the feature set was processed via 33 repetitions of 3-fold cross-validation (CV) of the training cohort with the machine learning framework FAMILIAR, using the MIFS algorithm for feature selection. A final radiomics signature was derived based on the power to classify resistant and sensitive tumors, using cumulative scoring across the CV folds and hyperparameter optimisation. A logistic regression model was trained to predict tumor radiosensitivity using the final signature. Its performance was validated on an independent data set.

Results:

Twenty-four tumors from one radioresistant model (SAS) and 24 tumors in total from two radiosensitive models (SKX and XF354) were used for the initial analysis. More than 250 ROIs per sensitivity class were considered and assigned to the training (2/3) and validation (1/3) cohorts. ROIs originating from the same tumor were treated as independent samples. 223 radiomic features were extracted from each ROI and the three best performing features for the classification of tumors as radioresistant or radiosensitive were identified (one morphological and two texture-based features). The logistic regression model using the final signature yielded a high accuracy: 0.96 (95% CI 0.94 – 0.98) and 0.93 (95% CI 0.89 – 0.97) for the training and validation cohort, respectively. Implementation of additional pre-processing and quality control steps such as stability analysis and batch effect control will be presented.

Conclusion:

Quantitative image features can be extracted from 2D preclinical immunofluorescence data with a potential to classify tumors based on their radiosensitivity. However, further modifications are required to increase the robustness of the classification signature. Radiomics analysis of preclinical image data can serve as a basis for biological hypotheses and design of preclinical validation experiments, and support the interpretability of clinically relevant radiomics models.