Document resource
Objectives We report on the clinical readiness of an AI tool for volumetric tumour quantification in pleural mesothelioma (PM) developed in the multicentre Mesothelioma detection and Quantification study, part of the PREDICT-Meso International Accelerator network.Methods We pre-specified a primary endpoint of ‘clinical acceptability’, requiring ≥90% of AI segmentations to be scored as ‘Excellent’ or ‘Satisfactory’ on a 5-point Likert scale and assembled a retrospective cohort of CT scans and clinical data, describing 1851 patients from 9 UK hospitals. 706/1851 had full metadata; 414/1851 had follow-up CT scans following therapy. 3 sites were held out for validation and to assess generalisation. Correlation and agreement (human edited vs. AI) were assessed using Dice overlap, Intra-Class Correlation (ICC) and Bland-Altman limits. AI- and human-segmented tumour volume (TV) were used to stratify patient response and compared using concordance index. Baseline AI-generated TVs were used to stratify overall survival using Cox proportional hazards models.Results Following full annotation of 353 scans (~422 CT slices per scan; >174,000 slices total), the primary objective was met. 90.8% of scans were deemed clinically acceptable in an external validation set of 98 scans from 9 different hospitals, with no performance difference in held-out scans, indicating good generalisation. 98% were considered ‘acceptable starting points’, needing no more than minor modifications to become clinically acceptable. The optimised network achieved Dice=0.95 and ICC=0.98, with non-significant mean bias of 28 cm3, suggesting edits did not meaningfully change the nature of the segmentation. Using mRECIST response cutoffs ([-30%,20%]), AI-derived TV outperformed human-segmented TV in 36 baseline-response CT pairs (concordance index=0.56 vs. 0.52). AI-derived TV stratified survival (414 paired CT scans) when dichotomised around the median volume (HR=0.68,p<0.001) in Cox proportional hazards models. Agreement on segmentation quality was perfect (κ=1). Clinical acceptability appeared independent of CT manufacturer and age (p=0.1844,p=0.1335).Conclusion These data demonstrate the clinical readiness of a fully-automated AI tool for PM segmentation. The optimised network generalises well to CT scans acquired on different scanners across multiple centres. AI-generated TV outperforms human-generated TV for response assessment and has important prognostic implications. Ongoing work includes tuning of optimal volumetric thresholds for response assessment.