ESTRO 2024 - Abstract Book

S5044

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

Material/Methods:

89 non-small cell lung cancer patients treated with 3D-CRT and IMRT who received a fractional size of 2.0–3.0 gray (Gy) / fraction (total 60–72 Gy) from Tohoku University Hospital were employed in this study. Whole lung CT ventilation images were used to calculate homodogy-based features. As detailed in Fig. 1, CT ventilation images were obtained by transforming 4DCT inhale and exhale images using deformable image registration and calculating from the change in HU of lung tissue between inhale and exhale. Three homology-based histograms were constructed for b0, b1 and ratio of b0 and b1 by setting the threshold levels ranging from 0 to 100. b0 and b1 are the Betti numbers and they represent the number of connected components and the number of holes respectively. From the histograms, 38 homology-based features are extracted. Homodogy-based features are obtained by multiplying the homology-based feature by dose-volume histogram (DVH) values, namely V 5 , V 20 , V 30 and MLD. Random Forest algorithm was adapted in this study and the prediction capability was compared using area under the ROC curve (AUC), accuracy, sensitivity and specificity values.

Fig. 1 Homodogy-based feature extraction and analysis workflow

Results:

Table I shows AUC, accuracy, sensitivity and specificity for the homodogy-based and radiomics feature. In general, the specificity scores were very low, and this may be attributable to the limited number of cases with RP. To improve accuracy and validate the result, larger cohort size is required for future research.

Table I. Results of homodogy-based features and radiomics features