ESTRO 2022 - Abstract Book

S1573

Abstract book

ESTRO 2022

R. Murphy 1 , N. Payan 1 , S. Osman 1,2 , K. Prise 1 , A. Hounsell 1,2 , J. O'Sullivan 1,3 , C. McGarry 1,4 , S. Jain 5,6

1 Queen's University Belfast, Patrick G. Johnston Centre for Cancer Research, Belfast, United Kingdom; 2 Belfast Health and Social Care Trust, Department of Radiotherapy Physics, Belfast, United Kingdom; 3 Belfast Health and Social Care Trust, Department of Clinical Oncology, Belfast, United Kingdom; 4 Department of Radiotherapy Physics, Belfast Health and Social Care Trust, Belfast, United Kingdom; 5 Queen's Unversity Belfast, Patrick G. Johnston Centre for Cancer Research, Belfast, United Kingdom; 6 Department of Clinical Oncology, Belfast Health and Social Care Trust, Belfast, United Kingdom Purpose or Objective More than 30% of prostate cancer patients will relapse within five years following radiotherapy treatment. Clinicopathological features do not have the ability to predict patients that will relapse and benefit from treatment intensification. Radiogenomics classification models have also demonstrated prognostic value in various cancer types. However, radiogenomics classification models in prostate cancer have yet to be developed. Our study aimed to investigate the prognostic ability of clinicopathological, radiomics, gene expression, and radiogenomics classification models for predicting disease progression in prostate cancer patients who have received radiotherapy plus androgen deprivation therapy (ADT). Materials and Methods 184 prostate cancer patients had their tumour transcriptome captured via microarray with 19453 features, and 5983 radiomics features were extracted from their CT scans before treatment. A feature selection/random forest pipeline, and generalized linear models (GLMs) with feature selection via the least absolute shrinkage and selection (LASSO) and ElasticNet, were used as classification models. Biochemical recurrence (BCR) and metastatic disease were used as classification labels, with 43 patients having had BCR, and 18 patients with metastatic disease. Stratified partitioning on these classification labels randomly assigned two thirds of the cohort into the training dataset, and one third into testing dataset. For the radiomics features, pairwise Pearson correlations were calculated, and highly correlated features were removed (r > 0.8). For the gene expression features, differentially expressed genes (DEGs; p-value < 0.01) were identified for both the classification labels to be used as the gene expression input data. Results 1429 robust radiomics features were identified following ICC, and 154 radiomics features were selected following pairwise correlation filtering to be used as radiomics input data. 195 DEGs were associated with BCR and 233 DEGs were associated with metastatic disease. The results of the clinicopathological, radiomics, gene expression, and radiogenomics models for both BCR and metastatic disease prediction are presented in Table 1 and Figure 1.

A combined radiogenomics model has the greatest prognostic ability for BCR, despite performing only slightly better than the gene expression model. For metastatic disease prediction, the radiomics model had the greatest prognostic ability.

Conclusion