Cone-beam computed tomography-based radiomics for tumour regression assessment during radiotherapy of non-small-cell...

During radiotherapy of non-small-cell lung cancers, changes in radiomics features extracted from cone-beam computed tomography images can serve as unique predictors. For adapting treatment plans or assessing early treatment responses, there is a pressing need to screen cone-beam computed tomography-derived radiomics features for predicting tumour regression volume.

Our study enrolled 58 patients with non-small cell lung cancer undergoing radiotherapy, comprising dataset1, and an additional 32 patients who underwent an adaptive treatment approach, forming dataset2. For dataset1 patients, radiomics features were extracted from both the planning computed tomography and the first treatment cone-beam computed tomography. For dataset2 patients, features were also extracted from the replanning computed tomography and the corresponding cone-beam computed tomography after 20 fractions of radiotherapy. By employing correlation coefficient and intraclass correlation coefficient analyses on planning computed tomography and first treatment cone-beam computed tomography features, we identified robust and reproducible features in dataset1 patients. Reproducible delta radiomics features were further selected to characterize feature changes and screen delta radiomics features for tumour regression in dataset2 patients. Lastly, predictive delta radiomics features were screened using least absolute shrinkage and selection operator regression.

Based on dataset1, we selected 34 radiomics features. From these, 16 delta radiomics features were further screened using dataset2. The changes observed in these features were consistent across both computed tomography and cone-beam computed tomography images. Specifically, nine features exhibited significant differences (p<0.05) while seven features remained relatively unchanged (p>0.05). Ultimately, four delta radiomics features were identified to classify tumour regression volume with a 30 % threshold for cone-beam computed tomography images.

Our study introduces a novel strategy for selecting computed tomography-based cone-beam computed tomography radiomics features, enabling the identification of reliable features for model development. Preliminary findings firstly demonstrate the feasibility of using cone-beam computed tomography-based delta radiomics features to classify tumour regression volume during radiotherapy of non-small-cell lung cancer.