Combining MRI radiomics, hypoxia gene signature score and clinical variables for prediction of biochemical recurrence-free survival after radiotherapy in prostate cancer

Abstract

Purpose: To investigate the value of combining MRI radiomic and hypoxia-associated gene signature information with clinical data for predicting biochemical recurrence-free survival (BCRFS) after radiotherapy for prostate cancer.

Methods: Patients with biopsy-proven prostate cancer, hypoxia-associated gene signature scores and pre-treatment MRI who received radiotherapy between 01/12/2007 and 31/08/2013 at two cancer centres were included in this retrospective cohort analysis. Prostate segmentation was performed on axial T2-weighted sequences using RayStation (v9.1). Histogram standardisation was applied prior to radiomic feature (RF) extraction. PyRadiomics (v3.0.1) was used to extract RFs for analysis. Four multivariable Cox proportional hazards BCRFS prediction models using clinical information alone and in combination with RFs and/or hypoxia scores were evaluated using concordance index (C-index) [confidence intervals (CI)]. Akaike Information Criterion (AIC) was used to assess model fit.

Results: 178 patients were included. The clinical-only model performance C-index score was 0.69 [0.64-0.7]. The combined clinical-radiomics model (C-index 0.70[0.66-0.73]) and clinical-radiomics-hypoxia model (C-index 0.70[0.65-0.73]) both had higher model performance. The clinical-hypoxia model (C-index 0.68 [0.63-0.7) had lower model performance. Based on AIC, addition of RFs to clinical variables alone improved model performance (p = 0.027), whereas adding hypoxia gene signature scores did not (p = 0.625). The selected features of the combined clinical-radiomics model included age, ISUP grade, tumour stage, and wavelet-derived grey level co-occurrence matrix (GLCM) RFs.

Conclusion: Adding pre-treatment prostate MRI-derived radiomic features to a clinical model improves accuracy of predicting BCRFS after prostate radiotherapy, however addition of hypoxia gene signatures does not improve model accuracy.

Keywords: Hypoxia; Magnetic resonance imaging; Prostate cancer; Radiomics; Radiotherapy.

Fig. 1

Flowchart showing study pipeline from image segmentation, image normalisation, radiomic feature extraction, image post-processing, feature selection steps to model building integrating hypoxia and radiomic data with clinical data. COMBAT, combating batch effects when combining batches; MRMR, minimum redundancy maximum relevance

Fig. 2

Bar chart showing the frequency (%) that each radiomic feature was selected across all cross-validation runs (out of 200) for each feature selection method. LHH, HHL, HLH, LLL, 3D wavelet radiomic features; GLRLM, grey level run length matrix; GLCM, grey level co-occurrence matrix; GLDM, grey level dependence matrix; GLSZM, grey level size zone matrix

Fig. 3

C-index and confidence interval (CI) of all 4 models, showing the joint best models were Clinical + Radiomics (0.7) and Clinical + Radiomics + Hypoxia (0.7)