Development and evaluation of machine-learning methods in whole-body magnetic resonance imaging with diffusion weighted imaging for staging of patients with cancer: the MALIBO diagnostic test accuracy study

• DOI: 10.3310/kpwq4208
• Publication Date: 2024-10-14T13:19:48Z
• Authors: Andrea Rockall; Xingfeng Li; Nicholas Johnson; Ioannis Lavdas; Shalini Santhakum...; A Toby Prevost; Dow-Mu Koh; Shonit Punwani; Vicky Goh; Nishat Bharwani; Amandeep Sandhu; Harbir Sidhu; Andrew Plumb; James Burn; Aisling Fagan; Alf Oliver; Georg J Wengert; Daniel Rueckert; Eric Aboagye; Stuart A Taylor; Ben Glocker
• Keywords: Diffusion imaging

Background Whole-body magnetic resonance imaging is accurate, efficient and cost-effective for cancer staging. Machine learning may support radiologists reading whole-body imaging. Objectives To develop a machine-learning algorithm to detect normal organs lesions. compare diagnostic accuracy, time agreement of radiology reads metastases using with concurrent machine (whole-body + learning) against standard deviation). Design participants Retrospective analysis (1) prospective single-centre study in healthy volunteers > 18 years ( n = 51) (2) multicentre STREAMLINE patient data 438). Tests Index: learning. Comparator: deviation. Reference Previously established expert panel consensus reference at 12 months from diagnosis. Outcome measures Primary: difference per-patient specificity between Secondary: sensitivity, per-lesion sensitivity specificity, read agreement. Methods Phase 1: classification forests, convolutional neural networks, multi-atlas approaches organ segmentation. 2/3: scans were allocated 2 (training 226, validation 45) 3 (testing 193). Disease sites manually labelled. The final was applied 193 cases, generating probability heatmaps. Twenty-five (18 experienced, 7 inexperienced imaging) randomly or deviation over two three rounds National Health Service setting. Read independently recorded. Results Phases 1 2: network had best Dice similarity coefficient, recall precision measurements Final used ‘two-stage’ initial identification followed by lesion detection. 3: evaluable (188/193, which 50 117 colon, 71 lung cases) November 2019 March 2020. For experienced readers, detection 86.2% 87.7% deviation), (difference −1.5%, 95% confidence interval −6.4% 3.5%; p 0.387); 66.0% 70.0% deviation) −4.0%, −13.5% 5.5%; 0.344). readers (53 reads, 15 metastases), 76.3% both groups sensitivities 73.3% 60.0% Per-site remained high within all sites; above (experienced) 90% (inexperienced). highly variable due low number lesions each site. Reading lowered under 6.2% (95% −22.8% 10.0%). primarily influenced round times reduced 32% 20.8% 42.8%) overall subsequent regression showing significant effect 0.0281) estimated as 286 seconds (or 11%) quicker. Interobserver variance suggests moderate agreement, Cohen’s κ 0.64, 0.47 0.81 0.66, Limitations Patient heterogeneous relatively few metastatic wide variety locations, making training testing difficult hampering evaluation sensitivity. Conclusions There no accuracy without support, although be slightly shortened Future work Failure-case improve model training, automate segmentation transfer techniques other tumour types modalities. Study registration This registered ISRCTN23068310. Funding award funded the Institute Care Research (NIHR) Efficacy Mechanism Evaluation (EME) programme (NIHR ref: 13/122/01) published full ; Vol. 11, No. 15. See NIHR Awards website further information.