P10.17.B IMPACT OF INTENSITY STANDARDISATION AND COMBAT BATCH SIZE VARIATION ON COMBINED CLINICAL AND RADIOMICS PROGNOSTIC MODELS IN PATIENTS WITH UNIFOCAL GLIOBLASTOMA

• Published in: Neuro-oncology (Charlottesville, Va.) Vol. 26; no. Supplement_5; p. v59
• Main Authors: Fatania, K, Frood, R, Mistry, H, Short, S C, O’Connor, J, Scarsbrook, A F, Currie, S
• Format: Journal Article
• Language: English
• Published: 17-10-2024
• ISSN: 1522-8517; 1523-5866
• DOI: 10.1093/neuonc/noae144.193

Abstract BACKGROUND Prognostic models for patients with glioblastoma (GBM) may include quantitative image features extracted from MRI (‘radiomics’). Multicentre image acquisition poses a challenge to radiomic feature (RF) reproducibility and MRI intensity standardisation techniques (ISTs) and ComBat may help. Optimal IST and minimum batch size (MBS) for ComBat are uncertain - the latter especially with real-world datasets where few patients are scanned at each site. Our aim was to assess the effect of three popular ISTs in conjunction with varying MBS on prognostic model performance and stability in a real-world, multi-centre GBM patient cohort. MATERIAL AND METHODS Patients with IDH-wildtype unifocal WHO grade 4 GBM and pre-operative MRI acquired across multiple geographic locations in the region between 2014-2020 were retrospectively included. Tumours were segmented using deep-learning with manual correction. WhiteStripe (WS), Nyul histogram matching (HM) and Z-score (ZS) ISTs were applied before RF extraction. Post-extraction, RFs were realigned using ComBat using MBS of 5, 10 or 15. Cox proportional hazards models were produced using five strategies to select the four best performing RFs, and the impact of IST and MBS was evaluated with bootstrapping (1,000 repetitions), and measuring model calibration, discrimination, relative explained variation and model fit. Model instability was assessed using bootstrap resampling to produce 95% confidence intervals (95% CIs), feature instability tables and calibration instability plots. RESULTS 195 patients were included with median overall survival of 13 months (95% CI 12-14). Increasing the ComBat MBS reduced the sample size available for modelling (195, MBS=5; 159 MBS=10; 129 MBS = 15). HM and WS tended to produce the highest relative increase in model discrimination, explained variation and model fit but IST choice did not influence model stability or calibration. Increasing the ComBat MBS improved model discrimination, relative explained variance and model fit, however it degraded model stability and calibration accuracy. Calibration instability plots showed a greater spread of calibration curves across the bootstrap resamples as the ComBat MBS increased (and sample size reduced), indicating worse calibration accuracy and stability. CONCLUSION IST and ComBat MBS had varied effects. Increased ComBat MBS (and reduced sample size) improved most performance metrics but stability and calibration deteriorated. HM and WS tended to increase most model performance metrics and did not negatively influence stability. Our real-world, multi-centre data showed that while ComBat may improve multi-centre radiomics models, it requires data to be excluded, with negative results for sample size and stability and unknown consequences for patients imaged in smaller centres.