Davies multicentre MRS study · – of prognosis to improve treatment stratification, ... • Data collection issues – Non-standard data formats for MRS raw data complicates data

Experiences of an ongoing multicentre study

investigating the use of 1H MRS to aid the clinical

management of childhood brain tumours

Davies NP, Arvanitis TN, Auer D, French A, Grazier R, Grundy R, Hargraves D, Howe FA, Jaspan T, Lateef S, Leach MO, MacPherson L, Natarajan K, Orphanidou E, Payne

G, Saunders D, Sun Y, Peet AC

Background

• Cancer: commonest cause of death from

disease in children.

• MRI techniques probing tumour biology provide

potential non-invasive biomarkers and surrogate

markers of response to targeted agents.

• Small number of children treated in one centre

⇒ Robust evaluation of imaging biomarkers can only take place in a multi-centre setting.

Participating Centres

• University of Birmingham / Birmingham Children’s Hospital

• Institute of Cancer Research / Royal Marsden Hospital

• University of Nottingham / Queen’s Medical Centre

• University College London / Great Ormond Street Hospital

• St George’s University London

• Alderhey Children’s Hospital, Liverpool

Programme Development

• Functional Imaging Group of Childrens Cancer and Leukemia Group (CCLG)

• MRS methods for characterisation of childhood brain tumours (5+ years collaboration)

• CRUK / EPSRC 5 year programme combining MRS with diffusion / perfusion MRI applied to specific scientific hypotheses and management questions in childhood cancer

Aims and Objectives

• To develop functional imaging techniques in parallel with

novel molecularly targeted agents in collaboration with the

CCLG Therapeutics Steering Group and Biological Studies

Group.

• To develop and evaluate imaging biomarkers:

– for improved non-invasive classification

– of prognosis to improve treatment stratification,

– for treatment planning

– predictive of drug response to tailor treatment to the individual,

– for early monitoring of drug response to optimise treatment.

Database Development

• Robust database design

• Clinical, surgical, histological and MRI/MRS data

• Auto archiving, complete log of any changes

• Easy to use remote data entry systems

– PHP based web application

– Index tables to tie interface to schema

• Allow for future data storage requirements

• Platform independent software

CCLG Database Structure

Patient

Trial Study

Histology / Diagnosis

Presentation

MRI

MRS

Surgery

Web application GUI GUI to DB Index Table Database

Hospitals

Oncologists

WHO diagnosis

Other lists

(Yes/No/Unknown/etc)

Web to Database link:

Users

Remote Data Entry

QA: Localisation Phantom

Inner Cube: (pH 7.6)

0.1 M Li lactate

0.1 M Cr

2cm

2cm

15cm

0.3cm

2cm

21cm

Outer Volume: (pH 8.3)

0.15 M Na acetate

QA Results: Localisation

Siemens 1.5T Siemens 1.5T

GE 1.5TPhilips 1.5T

Accrual

• Over 320 cases from 6 centres

– Approx 200 from BCH + 120 from other centres

– Follow-up MRS available in many cases

• Data collection issues

– Non-standard data formats for MRS raw data complicates

data capture and transfer

– Incomplete datasets (lack of clinical info, MRS raw data)

– Adherance to protocol not 100%

– Growing prevalence of 3 T scanners

– Compatibility of MRS?

Study 1: Brain Stem Tumours

• Important clinical group due to difficulties of surgery and often very poor prognosis

– Diagnosis often by radiological / clinical criteria

• Accrual: 32 cases from 4 centres

– Central radiological review

– Diagnostic categories:

• Diffuse pontine glioma (DPG) / diffuse glioma (DG)

• Focal low grade glioma (LGG)

• Focal high grade glioma (HGG)

• Uncertain

Study 1: BST Results

• DPG+DG had higher Cr / Cho (P<0.01) and mI / Cho

(P<0.01) ratios vs focal tumours

• LGG had higher mI / Cho vs HGG (P<0.05)

• DGs had higher lipid levels vs DPGs (P<0.05).

• Lipid levels in DGs were comparable to focal HGG

(higher than LGG)

Study 2: Low Grade Gliomas

• Background

– LGG: 40-50% of CNS tumours

– Diverse group, many follow more indolent course but

some have poorer prognosis

– Many decision points for treatment

– Need for non-invasive aids to clinical decision-making

• Aims

– Develop non-invasive biomarkers for characterisation and

monitoring of LGG in childhood using 1H MRS

Study 2: LGG Accrual

• Challenge of collating and validating more detailed

clinical information = 2 centres so far

• Total of 110 patients with pre-treatment MRS

– 34 patients failed QC, 12 patients had missing data.

– 64 patients for further analysis (57 BCH, 7 QMC)

• Many diagnostic categories and

sub-categories

LGG Mean Spectra

Pilocytic

Astrocytoma

Unbiopsied optic

pathway glioma

Diffuse

astrocytoma

Tectal plate

glioma

DNET Ganglioglioma

Pilocytic Astrocytomas

Linear Discriminant Analysis of 3 PA sub-groups

Study 2: Summary

• Significant metabolic differences found between LGG based on location of tumour, NF1 status and metastatic status

– Diagnostic aid

• Promising prognostic indicators found for sub-groups of LGG

– Potential guidance for treatment decisions and monitoring

Study 3: Prospective Classification

• Purpose

– To evaluate MRS as a diagnostic tool for paediatric

brain tumours

• Method

– Use multi-centre MRS data as an independent,

“pseudo-prospective” test of classifiers trained

with single-centre MRS data

Training dataset:

* acquired Mar 2003 – Apr 2008 with biopsy

MRS acquisition:

• Following standard pre- & post-contrast MRI

• PRESS (TR 1500ms / TE 30ms), Bandwidth 1 Hz/point

• Single-Voxel placed inside solid/enhancing portion

• Voxel vol. 3.4 ml (NSA 256) – 8 ml (NSA 128)

Methods: Classifier Training

Centre No. of test cases Scanners

BCH 83 * Siemens & GE 1.5 T

Methods: Classifier Training

Classifier Methodology: (Davies et al NMR Biomed 2008)

• MRS processing and metabolite profile estimation: LCModel™(Provencher, Magn Reson Med 1993)

• QC filter: SNR > 4; Linewidth(H2O) < 10 Hz; artefacts

• Principle Components Analysis (PCA)

• Linear Discriminant Analysis (LDA)

• Cross-validation (Leave-one-out and Bootstrap)

• Two binary classifiers: (Davies et al, ESMRMB 2008)

1) PNET vs glial

2) Low Grade (WHO I&II) vs High Grade (WHO III & IV)

Test dataset:

* May08–Dec09 / * QC + diagnosis / * QC + biopsy

* 1.5 or 3T, TE (23-40ms), TR (1500-2000ms)

• QC filter: same as for training

Methods: Classifier Testing

Centre No. of test cases Scanners

BCH 49* / 39* / 33* Siemens & GE 1.5 T

QMC 35 * Philips 1.5 T & 3 T

GOS 8 Siemens 1.5 T

RMH / SGH 10 Philips 1.5 T

Combined 92*

Glial vs PNET Classifier

Results: Prospective Classifier

Test cohort

Classifier Accuracy

PNET vs Glial

Low Grade vs High

Grade

BCH 90% (N=31) 87% (N=39)

External Centres 79% (N=42) 81% (N=37)

Combined 84% (N=73) 85% (N=76)

Conclusions

• Multicentre studies involving 1H MRS are possible and

can be clinically useful, but challenges remain

• Multivariate analysis of MRS is a promising tool for

enhanced clinical management of childhood brain

tumours

• Further development of multicentre MRS data

collection and analysis is justified

– Combination with perfusion and diffusion MRI

Acknowledgements

• All members of the CCLG Functional Imaging

Group

• Radiographers and data managers in

participating centres

• Funding bodies:

– CRUK, EPSRC, MRC, NIHR, eTUMOUR