ESTRO 2024 - Abstract Book

S5152

Physics - Radiomics, functional and biological imaging and outcome prediction

ESTRO 2024

minimum redundancy and bootstrap-based machine learning approach to select the most significant RF and develop models. The resulting Prognostic Index (PI) was then tested on the validation cohort.

Results:

222 LARC patients were included (150 training and 72 validation cohorts).

54 (24.3%) achieved ‘complete response’. The first model variable included was Morphology_Diameter with a negative b coefficient, (b=-0.039377) meaning lower diameter tumours associated with 'complete response'. The second variable was the cm_inv_diff_norm_3D_avg which measures Grey-level homogeneity. The feature score is maximal if all Grey-levels are the same. In this scenario the variable was found with a negative b coefficient, unveiling that lower values of cm_inv_diff_norm_3D_avg were associated with ‘complete response’. When these two RFs were combined to form a PI, the model associated with ‘complete response’ ( AUC=0.70, 95%CL 0.62-0.78, p= 0.0001 ) in training and the results were confirmed in validation (A UC=0.70, 95%CI 0.60-0.80, p=0.0242 )(Fig1). On analysis of tumour regression, 139 (63%) achieved ‘good regression’. The first variable included in the regression model was the morph_vol_approx with a positive b coefficient (b=0.0000092075), meaning higher volume of lesion associated to ‘poor regression’. The second feature was szm_szhge_3D . According to IBSI, this emphasizes zone counts in the lower left quadrant of the Gray-Level-Size-Zone-Matrix, where small zone sizes and high Grey-levels are located. In the model, the features were found with positive b coefficient meaning that higher values of the feature are likely associated to poor regression. On this basis, the PI from these combined RFs associated with poor regression in training ( AUC=0.64, 95%CI 0.56-0.72, p= 0.0039 ) and validation ( AUC=0.66, 95%CI 0.53-0.84, p=0.0398 )(Fig1).

Conclusion: