Machine Learning in Myeloma Response (MALIMAR)

Development of a Machine Learning Support for Reading Whole Body Diffusion Weighted Magnetic Resonance Imaging (WB-DW-MRI) in Myeloma for the Detection and Quantification of the Extent of Disease Before and After Treatment

Diffusion-weighted Whole Body Magnetic Resonance Imaging (WB-MRI) is a new technique that builds on existing Magnetic Resonance Imaging (MRI) technology. It uses the movement of water molecules in human tissue to define with great accuracy cancerous cells from normal cells. Using this technique the investigators can much more accurately define the spread and rate of cancer growth. This information is vital in the selection of patients' treatment pathways. WB-MRI images are obtained for the entire body in a single scan. Unlike other imaging techniques such as computed Tomography (CT) or Positron Emission Tomography (PET) PET/CT there is no radiation exposure. Despite the considerable advantages that this new technique brings, including "at a glance" assessment of the extent of disease status, WB-MRI requires a significant increase in the time required to interpret one scan. This is because one whole body scan typically comprises several thousand images. Machine learning (ML) is a computer technique in which computers can be 'trained' to rapidly pin-point sites of disease and thus aid the radiologist's expert interpretation. If, as the investigators believe, this technique will help the radiologist to interpret scans of patients with myeloma more accurately and quickly, it could be more widely adopted by the NHS and benefit patient care. The investigators will conduct a three-phase research plan in which ML software will be developed and tested with the aim of achieving more rapid and accurate interpretation of WB-MRI scans in myeloma patients.

Rationale: Diffusion-weighted whole body magnetic resonance imaging (WB-MRI) is a technique that depicts myeloma deposits in the bone marrow. WB-MRI covers the entire body during the course of a single scan and can be used to detect sites of disease without using ionising radiation. Although WB-MRI allows for "at a glance" assessment of disease burden, it requires significant expertise to accurately identify and quantify active myeloma. The technique is time-consuming to report due to the great number of images. A further challenge is recognising whether a patient has residual disease after treatment. Machine learning (ML) is a computer technique that can be trained to automatically detect disease sites in order to support the radiologist's interpretation. The investigators believe this technique will help the radiologist to interpret the scan more accurately and quickly. Machine learning algorithms have been successfully developed to recognise some other cancer types. The investigators believe that it may be successful in patients with myeloma, in whom The National Institute for Health and Care Excellence (NICE) recommend whole body MRI. This could allow the technique to be more widely used in the National Health Service (NHS). In the MALIMAR study the investigators will develop and test ML methods that have the potential to increase accuracy and reduce reading time of WB-MRI scans in myeloma patients. The investigators propose to develop ML tools to detect and quantify active disease before and after treatment based on WB-MRI. Research will be carried out at the Royal Marsden Hospital (RMH) NHS Foundation Trust, Institute of Cancer Research (ICR) London and Imperial College London. The investigators will use Whole Body MRI (WB-MRI) scans that have already been acquired in myeloma patients. They will also include 50 new scans obtained at RMH from healthy volunteer scans which will be used to 'teach' the computer to distinguish between healthy and diseased tissues. Research Design: The research will be divided into three parts: Development of the Machine Learning (ML) tool to detect active myeloma Measurement of the ability of the ML tool to improve the radiologists' interpretation of WB-MRI scans using a set of scans from patients with active and inactive myeloma and new scans obtained from healthy volunteers Development of the ML tool to quantify disease burden and changes between pre- and post-treatment WB-MRI scans in order to identify response to treatment The main outcome measure for this study will be the improvement in the detection of active disease and disease burden and the reduction in radiology reading time. The investigators will assess the reduction in reading time in both experienced specialist and non-specialist radiologists.

Whole Body Diffusion Weighted, Magnetic Resonance Imaging, Machine Learning, Reading Time, Convolutional Neural Network, Algorithm, Diagnostic performance

Cross-sectional diagnostic test accuracy design: development of a machine-based algorithm to augment expert classification of disease status and response to treatment in myeloma patients using retrospective interpretation of WB-MRI scans and disease-free (healthy volunteers) for comparison purposes.

The assessors interpretation of disease status using WB-MRI scans will be fully blinded to the reference standard (i.e. the Expert Panel's interpretation of the same scan).

Machine learning (ML): A mixed data set of 200 WB-MRI scans comprising scans obtained from 40 healthy volunteers (scanned for the purposes of the study), 40 previously acquired inactive myeloma WB-MRI scans and 120 previously acquired active myeloma WB-MRI scans, in which machine learning and convolutional neural networks will be trained to recognise healthy marrow, treated inactive previous myeloma and active myeloma. An algorithm will be developed for testing in phase 2.

Machine Learning (ML): A mixed data set of 353 WB-MRI scans as that comprising 50 healthy volunteers (scanned for the purposes of the study), and previously acquired scans from 303 myeloma patients, 100 of whom have inactive disease and 203 of whom have active myeloma. The scans will be read by radiologists in random order either with or without the support of for the detection of active myeloma. The diagnostic performance of the radiology reads with or without the machine learning support will be measured against an expert panel reference standard.

Machine Learning (ML): Approximately 200 paired WB-MRI scans from 100 patients (scanned at baseline with active disease and then post treatment) will be used to develop a machine learning tool to quantify the burden of disease. The machine learning algorithm will then be tested on a further additional set of 60 patients who previously had two WB-MRI scans comprising paired baseline (with active disease) and post treatment scans. The agreement of radiology readers to evaluate the burden of disease will be measured against the reference standard (expert panel) with and without machine learning support.

Application of ML support algorithm to accelerate and enhance human interpretation of WB-MRI scans in patients with myeloma

Sensitivity for the detection of active myeloma on WB-MRI with and without ML support versus the reference standard

Agreement between readers and reference standard in scoring overall disease burden with and without ML intervention

Agreement of machine learning algorithm with reference standard to classify disease spread assessed as percentage accuracy

Inclusion Criteria (healthy volunteers): Able to provide written informed consent No contra-indication to MRI 40 years or above in age (age matched as far as possible to WB-MRI scan set) No known significant illness No known metallic implant Exclusion Criteria: Not able to provide written informed consent A contra-indication to MRI <40 years or above in age (age matched as far as possible to WB-MRI scan set) A known significant illness A known metallic implant

Satchwell L, Wedlake L, Greenlay E, Li X, Messiou C, Glocker B, Barwick T, Barfoot T, Doran S, Leach MO, Koh DM, Kaiser M, Winzeck S, Qaiser T, Aboagye E, Rockall A. Development of machine learning support for reading whole body diffusion-weighted MRI (WB-MRI) in myeloma for the detection and quantification of the extent of disease before and after treatment (MALIMAR): protocol for a cross-sectional diagnostic test accuracy study. BMJ Open. 2022 Oct 5;12(10):e067140. doi: 10.1136/bmjopen-2022-067140.