Neoadjuvant chemotherapy in breast cancer: early response prediction with quantitative MR imaging and spectroscopy.

Abstract

A prospective study was undertaken in women undergoing neoadjuvant chemotherapy for locally advanced breast cancer in order to determine the ability of quantitative magnetic resonance imaging (MRI) and proton spectroscopy (MRS) to predict ultimate tumour response (percentage decrease in volume) or to detect early response. Magnetic resonance imaging and MRS were carried out before treatment and after the second of six treatment cycles. Pharmacokinetic parameters were derived from T1-weighted dynamic contrast-enhanced MRI, water apparent diffusion coefficient (ADC) was measured, and tissue water:fat peak area ratios and water T2 were measured using unsuppressed one-dimensional proton spectroscopic imaging (30 and 135 ms echo times). Pharmacokinetic parameters and ADC did not detect early response; however, early changes in water:fat ratios and water T2 (after cycle two) demonstrated substantial prognostic efficacy. Larger decreases in water T2 accurately predicted final volume response in 69% of cases (11/16) while maintaining 100% specificity and positive predictive value. Small/absent decreases in water:fat ratios accurately predicted final volume non-response in 50% of cases (3/6) while maintaining 100% sensitivity and negative predictive value. This level of accuracy might permit clinical application where early, accurate prediction of non-response would permit an early change to second-line treatment, thus sparing patients unnecessary toxicity, psychological morbidity and delay of initiation of effective treatment.

Full text

Page 1

Neoadjuvant chemotherapy in breast cancer: early response
prediction with quantitative MR imaging and spectroscopy
DJ Manton*,1, A Chaturvedi2, A Hubbard3, MJ Lind1, M Lowry1, A Maraveyas1, MD Pickles1, DJ Tozer4and
LW Turnbull1
1The Postgraduate Medical Institute of the University of Hull Division of Cancer in association with the Hull York Medical School, The University of Hull,
Hull, East Yorkshire HU6 7RX, UK;2The Department of Clinical Oncology, Hull and East Yorkshire Hospitals NHS Trust, The Princess Royal Hospital,
Saltshouse Road, Hull, East Yorkshire HU8 9HE, UK;3The Breast Screening Unit, Hull and East Yorkshire Hospitals NHS Trust, Castle Hill Hospital,
Castle Road, Cottingham, East Yorkshire HU16 5JQ, UK;4NMR Research Unit, Institute of Neurology, Queen Square, London WC1N 3BG, UK
A prospective study was undertaken in women undergoing neoadjuvant chemotherapy for locally advanced breast cancer in order to
determine the ability of quantitative magnetic resonance imaging (MRI) and proton spectroscopy (MRS) to predict ultimate tumour
response (percentage decrease in volume) or to detect early response. Magnetic resonance imaging and MRS were carried out
before treatment and after the second of six treatment cycles. Pharmacokinetic parameters were derived from T1-weighted dynamic
contrast-enhanced MRI, water apparent diffusion coefficient (ADC) was measured, and tissue water:fat peak area ratios and water T2
were measured using unsuppressed one-dimensional proton spectroscopic imaging (30 and 135ms echo times). Pharmacokinetic
parameters and ADC did not detect early response; however, early changes in water:fat ratios and water T2(after cycle two)
demonstrated substantial prognostic efficacy. Larger decreases in water T2accurately predicted final volume response in 69% of cases
(11/16) while maintaining 100% specificity and positive predictive value. Small/absent decreases in water:fat ratios accurately
predicted final volume non-response in 50% of cases (3/6) while maintaining 100% sensitivity and negative predictive value. This level
of accuracy might permit clinical application where early, accurate prediction of non-response would permit an early change to
second-line treatment, thus sparing patients unnecessary toxicity, psychological morbidity and delay of initiation of effective treatment.
British Journal of Cancer (2006) 94, 427–435. doi:10.1038/sj.bjc.6602948
& 2006 Cancer Research UK
www.bjcancer.com
Keywords: breast neoplasms; chemotherapy; magnetic resonance; MR; spectroscopy; contrast enhancement; diffusion study
? ???????????????????????????????????????????????????
Recent statistics confirm that breast cancer remains one of the
most prevalent and serious forms of neoplastic disease in the UK.
Breast cancer accounts for more than one in four female cancer
cases and currently demonstrates age-standardised incidence and
mortality rates of 117 and 31 per 100000, respectively (Toms,
2003).
Some primary breast cancers are considered inoperable at
diagnosis owing to their problematical location and size, with
tumours greater than 3cm in diameter being associated with an
increased risk of disseminated disease. In such cases, neoadjuvant
chemotherapy is routinely used before surgery to increase the
chances of a successful outcome. Assessing tumour response to
chemotherapy is crucial to patient management and is currently
achieved by monitoring changes in tumour size using clinical
examination backed up by longitudinal X-ray or ultrasound
mammography. A poor response to the primary treatment regime
usually prompts either change of chemotherapy regime or an early
resort to surgery. Poor response might also require a greater
degree of surgical intervention.
Repeated X-ray mammography (XRM) has several drawbacks:
discomfort, radiation exposure, geometric distortion owing to
compression and magnification render tumour volume measure-
ments less accurate than magnetic resonance imaging (MRI), and
tumour may be impossible to distinguish from dense glandular
tissue and fibrosis. Magnetic resonance imaging has been shown to
correspond better with pathological size measurement than XRM
(Esserman et al, 1999; Drew et al, 2001) and to be less obscured by
dense breast diseases. Ultrasound measurement is more accurate
than XRM, but fails when the tumour to be measured is larger than
the field of view or complex in shape. Magnetic resonance imaging
is also more accurate than ultrasound at detecting small volume
residual tumour.
However, even if tumour volume changes can be accurately
assessed by MRI, they may manifest themselves later than changes
in underlying tumour function such as vascular density or
permeability (Wasser et al, 2003). Therefore, vascular or metabolic
parameters might provide a more sensitive indicator of early
tumour response, thus permitting individual treatment regimes to
be adjusted more rapidly, and sparing patients unnecessary
morbidity, expense and delay in initiation of effective treatment.
Treatment-induced changes in metabolism can be investigated
using both positron emission tomography (Byrne et al, 2004;
Kumar and Alavi, 2004; Weber, 2005) and single photon emission
computed tomography (Buscombe et al, 1997), but the clinical
utility of such examinations is reduced by the need to limit
repeated radiation doses.
Treatment-induced changes in tumour neovasculature can be
assessed using dynamic contrast-enhanced MRI (DCE-MRI) and
Received 10 June 2005; revised 17 October 2005; accepted 7 December
2005
*Correspondence: Dr DJ Manton, MRI Centre, Hull Royal Infirmary,
Anlaby Road, Hull HU3 2JZ, UK; E-mail: d.j.manton@.hull.ac.uk
British Journal of Cancer (2006) 94, 427–435
& 2006 Cancer Research UKAll rights reserved 0007– 0920/06$30.00
www.bjcancer.com
Molecular Diagnostics

Page 2

quantified using pharmacokinetic (PK) modelling (Padhani, 2002;
Knopp et al, 2003). Diffusion-weighted MRI can be used to detect
changes in the apparent diffusion coefficient (ADC) for tissue
water associated with changes in tissue and intracellular structure
(Zhao et al, 1996; Gallons et al, 1999; Ross et al, 2003; Moffat et al,
2005). Proton magnetic resonance spectroscopy (MRS) can be used
to characterise breast lesions through differences in the ratio of fat
and water signals (Chu et al, 1987; Sijens et al, 1988) or the
intensity of signal from choline-containing compounds (Bradamante
et al, 1988; Katz-Brull et al, 2002; Bolan et al, 2003; Jacobs et al,
2004; Jacobs et al, 2005). Both fat:water ratios (Jagannathan et al,
1998) and choline levels (Preul et al, 2000; Jagannathan et al, 2001;
Schwarz et al, 2002; Meisamy et al, 2004) have also been used to
monitor treatment-induced changes, and phosphorous MRS
has been used to predict response using differences in the levels
of phosphocholine or phosphomonoesters (Shukla-Dave et al,
2002a,b; Arias-Mendoza et al, 2004). Sodium MRI might also be
useful in monitoring treatment response (Kline et al, 2000;
Schepkin et al, 2005).
A prospective study was, therefore, undertaken whereby three of
these techniques (DCE-MRI, ADC measurement and MRS fat:-
water ratios) were carried out in the same patients to determine in
each technique the relative prognostic utility. The study hypothesis
was that a combination of the quantitative data from the three
techniques would lead to a synergistic increase in prognostic
accuracy and, therefore, provide an accurate, reliable and non-
invasive early indication of tumour response to chemotherapy,
which could potentially have a positive influence on patient
management.
MATERIALS AND METHODS
Chemotherapy regime and relative timing of MRI
All women who were scheduled to undergo neoadjuvant chemo-
therapy for primary inoperable locally advanced breast cancer
were approached to join the study, which had received Local
Research Ethics Committee approval (reference number 03/00/
038). Longitudinal MRI was carried out using a 1.5T Signa
Advantage clinical MRI scanner (International General Electric,
Milwaukee, WI, USA) using a dedicated bilateral breast coil
(Machnet BV, Eelde, Netherlands). A standard dosage chemother-
apy regime was used involving intravenous administration of
epirubicin (bolus: 60mgm?2body surface area), cyclophospha-
mide (bolus: 600mgm?2) and 5-fluorouracil (continuous infusion
by pump: 200mgm?2per day). Good clinical practice, including
checking patients’ cardiac status before therapy if necessary, was
adhered to throughout the study. Magnetic resonance imaging and
MRS were carried out at three time points: before the first course
of chemotherapy (TP0), after the second course but no more than
55 days after the first course (TP2) and after the final (generally
sixth) course (TPF). Therapeutic response was confirmed after
surgery by pathological examination.
MR imaging and spectroscopy
All the MRI data at each time point were acquired in a single
imaging session with an average length of 1h. Dynamic contrast-
enhanced MRI was carried out first, during which gadopentetate
dimeglumine (formerly known as Gd-DTPA; Schering Health
Care, Burgess Hill, UK) was injected as a bolus at a dose of
0.1mmolkg?1body weight, using a two-dimensional, T1-weighted
multislice, fast RF spoiled gradient echo (FSPGR) sequence. In
order to attain adequate tissue coverage in all cases, imaging was
carried out in either the coronal (30cm field-of-view, FOV) or
sagittal (20cm FOV) planes (256?128 matrix in all cases) with
five, seven or nine slices (4–9mm thick, 2mm gaps). Scan timing
parameters, therefore, ranged as follows: TR¼8.8–11.1ms;
TE¼4.2ms fractional; flip¼301; temporal resolution¼10.5–
14.5s; 35 time points per slice; scan time¼6min 8s to 8min
28s. Dynamic contrast-enhanced MRI was preceded by proton
density-weightedFSPGRimaging
fractional/81) to enable correction for differences in native tissue
T1.
A single region-of-interest (ROI) that best delimited the lesion
present was drawn, by an experienced radiologist, for each DCE-
MRI study. In-house software (developed using the IDL language,
Research Systems Inc., Boulder, CO, USA) was then used to
measure the mean pixel signal intensity (SI) within the ROI for
the proton density image (SIPD) and each of the images in the T1-
weighted series (SIT1). These data were then used to calculate an
enhancement factor (EF) time series proportional to the concen-
tration of gadopentetate dimeglumine, with differences/changes
in the native tissue T1 between patients/visits being corrected
for using the proton density SI and the mean pre-contrast T1SI
as obtained from a user-defined number of baseline points
(SI0) (Hittmair et al, 1994). The equations for EF calculation
were EFðtÞ ¼ 1=ðKTRT1Þ?ln½ðSImax? SI0Þ=ðSImax? SIT1ðtÞÞ? and
SImax¼ SIPD?sinðaT1=aPDÞ, where aT1and TRT1are the flip angle
and TR for the dynamic T1-weighted series and aPDis the flip
angle for the proton density-weighted image. A value of 11.04 was
used for the constant K, this having been determined through
calibration experiments using gels of known T1values and being
valid for aT1¼ 30?only.
A two-compartment PK model (Buckley et al, 1994; Tofts, 1997)
was used to calculate the amplitude, or initial slope, of the EF time
series (this being proportional to the microvessel transfer constant
Ktrans(Tofts, 1997), itself proportional to the permeability surface
area product) and the rate constant (Kep). The model equation was
(TR/TE/flip¼120ms/4.2ms
EFðtÞ ¼Ktrans=ðKep? KelÞ
½expð?Kelðt ? t0ÞÞ ? expð?Kepðt ? t0ÞÞ?;
where Kel represents contrast clearance from the plasma and
t0represents the variable arrival time of the contrast agent. The
extracellular–extravascular tissue volume fraction (Ve) was also
estimated using the Ktrans/Kepratio (assuming equal influx and
efflux vascular permeability; Tofts, 1997). Curve fitting was carried
out (using mean ROI data, not pixel-by-pixel) using a nonlinear
least-squares algorithm implemented within the in-house, IDL
software, and the model-independent maximum enhancement
factor (MEF) was also recorded.
Tumour volume was measured using manually traced ROIs
drawn on high-resolution three-dimensional, post-contrast, fat-
suppressed FSPGR images. Again, sequence acquisition para-
meters were varied in order to obtain adequate tissue coverage in
all cases and these ranged as follows: TR/TE/flip¼14.6/4.2ms
fractional/301;plane/FOV/matrix¼sagittal/18–24cm/512?256;
slice number/thickness¼25/2.5–5mm. Tumour volumes were
measured, rather than more typical linear measurements, as this
was considered to be the most rigorous and accurate, although
most time-consuming, way of quantifying lesion size. Tumour
volume has been shown to correlate more closely than tumour
diameter with disease-free survival in breast cancer patients and it
may also provide a more sensitive characterisation of tumour
response than one-dimensional measurements (Partridge et al,
2005).
It was assumed that neoplastic tissue would not demonstrate
substantial ADC anisotropy; therefore, in an attempt to keep scan
time to a minimum, ADC was measured in the frequency-encode
direction only (left-to-right), which was known to demonstrate the
best imaging homogeneity and stability. The Stejskal–Tanner
technique (Stejskal and Tanner, 1965) was implemented using a
single fat suppressed slice (axial, 22cm FOV, 1282matrix, 7mm
thick) single-shot, spin-echo echo-planar imaging and with the
Predicting chemotherapy response with quantitative MRI
DJ Manton et al
428
British Journal of Cancer (2006) 94, 427–435
& 2006 Cancer Research UK
Molecular Diagnostics

Page 3

following diffusion weightings: 0, 14, 55, 125, 222, 346, 499 and
680smm?2. The ADC values were calculated on a whole ROI basis
(Gibbs et al, 2001), which included a correction for the bias
inherent in magnitude image reconstruction at low signal-to-noise
ratios (especially true for higher b-values) (Miller and Joseph,
1993).
The relative proportions of spectroscopic signal arising from
tissue water and lipids were measured at echo times of 30 and
135ms using a one-dimensional stimulated echo (STEAM)
spectroscopic imaging sequence (without chemical shift selective
suppression of any moieties) and a repetition time of 3s. Thirty-
two phase-encode steps were used over a 16cm superior–inferior
FOV with STEAM excitation limited to seven adjacent voxels
(1.0?0.5cm in the anterior–posterior and left–right directions,
respectively, giving a nominal voxel volume of 0.25ml). Figure 1
shows the seven spectroscopic voxels in a representative tumour
before chemotherapy. To ensure that the same region of tissue was
examined in follow-up studies, MRS voxels were relocated, by an
experienced radiologist, after careful comparison of the breast
architecture surrounding the lesion with the architecture demon-
strated on hard copies of previous voxel locations. Changes in
tumour choline were not measured, as the predictive utility of the
technique had not been established in the literature at the time of
study design.
Spectral analysis was carried out using the SAGE IDL package
(International General Electric, Milwaukee, WI, USA). In order to
decrease the degree of inter-voxel signal contamination, mild
spatial apodisation wasapplied
width¼90%/5%), which resulted in an effective voxel volume of
0.33ml (Jackson et al, 1994). To minimise partial volume
contamination from adipose tissues outside the tumours, only
signals from those voxels contained wholly within the tumour were
averaged; then, water and fat signal intensities were measured by
frequency-domain fitting of the water peak (4.7p.p.m.) and the
dominant lipid peaks at 1.3 and 0.9p.p.m. (representing the
(Fermifilter:diameter/
methylene –(CH2)n– and terminal methyl –CH3lipid moieties,
respectively). Spectroscopic water content (%WMRS) was then
quantified using the ratio of the water signal to that of the sum of
water and the two dominant fat signals. Water T2 was also
estimated using a crude two-point method: T2B(TE135?TE30)/
ln(area30/area135).
Biopsy data
Patients’ notes were reviewed to collate data obtained from
histopathological analysis of specimens acquired via pre-che-
motherapy core biopsy and final surgery. Parameters of interest
were histological tumour subtype, tumour grade (pre-chemother-
apy), the presence or absence of ductal carcinoma in situ (DCIS)
and pre-chemotherapy oestrogen and progesterone receptor status
(ERS and PRS) scores. Oestrogen and progesterone receptor status
scores were recorded in the notes on either a six- or eight-point
scale and converted to continuous data by calculating the ratio of
the score to maximum possible score. The maximum diameter of
the resected tumour, as calculated by the pathologist, was also
collated and compared to the maximum diameter recorded in the
final clinical MRI report using the Limits of Agreement method
proposed by Bland and Altman (1986).
Quantifying and correlating parameter changes and
tumour response
Changes in parameter values between TP0and TP2were calculated
either as absolute differences, D02ðXÞ ¼ XTP2? XTP0, or percentage
change, PC02ðXÞ ¼ XTP2=XTP0? 1, as deemed most appropriate. A
change in tumour volume at the end of treatment, PC0F(V),
exceeding ?65% was taken as the cutoff between partial response
(PR) and stable disease (SD), this being equivalent to a decrease
in cross-sectional area of 50%, itself broadly equivalent to the
RECIST criterion of a 50% decrease in the product of maximum
orthogonal diameters (Therasse et al, 2000). No attempt was made
to predict final tumour size in this study, as this is not likely to
alter patient management before surgery.
All statistical analyses were carried out using the SPSS package
(SPSS Inc., Chicago, IL, USA). Correlation between D02(X), PC02(X)
or continuous/ranked pathology data and PC0F(V) was assessed
using the Spearman non-parametric test (two-tailed). Correlation
between dichotomous pathology data (presence/absence of DCIS)
and categorical pathology data (cancer type) was assessed using
the Mann–Whitney and Kruskal–Wallis non-parametric tests
(two-tailed), respectively. Prognostic efficacy was assessed using
receiver–operator characteristic (ROC) curves (Altman and Bland,
1994), with PR as the positive result, and the areas contained
underneath them (AUC) calculated.
Variables were combined, in the hope of attaining a synergistic
increase in prognostic efficacy, using logistic regression analysis
(LRA) modelling, a statistical technique that maximises binary
classification accuracy using a linear combination of weighted
input variables plus a constant term (Hosmer and Lemeshow,
1989). Backwards conditional elimination of input variables
(P40.10) was used in order to prevent over-parameterisation of
models, and ordinal variables were treated as categorical with the
most benign category used as the indicator reference. Some
statistical confidence intervals (CIs) were obtained using tabulated
values (Anonymous, 1982).
RESULTS
Patient recruitment, withdrawal and other excluding
factors
A total of 46 women were recruited into the study over a period of
approximately 2 years; however, six (13%) of these subsequently
Figure 1
localisation technique in a pre-chemotherapy tumour. (The bottom two
voxels can be seen to contain wholly tumour, and spectral analysis would
have been limited to these voxels.) The image on which the voxels are
overlaid is the single slice, T1-weighted, sagittal localiser scan acquired with
fat suppression post-contrast and immediately before spectroscopy.
An example of the seven-voxel spectroscopic imaging
Predicting chemotherapy response with quantitative MRI
DJ Manton et al
429
British Journal of Cancer (2006) 94, 427–435
& 2006 Cancer Research UK
Molecular Diagnostics

Page 4

withdrew. One case of atypical, inflammatory lobular cancer was
excluded from data analysis on the grounds that it is impossible to
measure tumour volume accurately in such tumours. One case in
which histology only revealed DCIS was also excluded on the
grounds that DCIS is too dissimilar to the other tumour types
present, and chemotherapy is not the treatment of choice for it
(2/40¼5%).
In three of the remaining 38 cases, the second MRI examination
was severely delayed owing to clinical complications and was
carried out after the chosen cutoff of 55 days, thus necessitating the
exclusion of the TP2(but not the TP0) data. In addition, it was not
possible to acquire ADC and MRS data on every occasion because
of a number of problems including hardware failure and an
inability to acquire sufficient signal to permit sequence optimisa-
tion. Such failures occurred during the TP0scan, necessitating
complete exclusion of the patient from data analysis, in four cases,
and at TP2, necessitating removal of TP2data only, in nine cases. A
total of 13 failures in the 73 examinations (38 at TP0and 35 at TP2)
represent a failure rate of 18%. Results will, therefore, be presented
for 22 cases for which both TP0and TP2data were available and an
additional 12 cases for which data were available at TP0only (see
Figure 2).
Neither tumour volume measurements nor PK analysis were
affected at any time point, as these were based on three-
dimensional and DCE-MRI images acquired before ADC and
MRS were attempted. Therefore, tumour volume response,
PC0F(V), could be quantified in all cases. Pharmacokinetic data
and tumour volume data calculated in cases where ADC and MRS
data were not acquired were voluntarily excluded, so as not to bias
statistical power in favour of PK/volume parameters. In the 22
cases for which valid TP2data were available, 18 (82%) of the
women had received two courses of chemotherapy at the time of
MRI and four had received three courses. In three cases, it was
decided to cease chemotherapy before a full six cycles were
completed, as poor clinical response was being demonstrated. In
these cases, MRI was, however, carried out after the final course of
treatment, thus permitting a final measurement of tumour volume
to be made and, thus, quantifying the relatively poor response
adequately.
All five pathology variables (i.e. cancer type, grade, presence/
absence of DCIS, and ERS and PRS scores) were available in 35 of
the 38 eligible cases. These included 32 of the 34 valid TP0MRI
cases, all 22 of the valid TP2MRI cases and three cases in which
MRI data were unusable (see Figure 2). The age at TP0for the 37
women whose data are used in this study ranged from 26 years, 6
months to 75 years, 7 months, with a median of 48 years, 6 months.
The pathological cancer subtypes of these 37 cases were 19 ductal/
not specified (12 with DCIS), 15 (nine) ductal and three (one)
lobular. The number of women with cancers of grade one, two and
three were nine, 18 and eight, respectively (cancer grade was not
reported by the pathologist in two cases). Two women did not
undergo surgery because of the presence of metastatic disease and
one woman did not undergo surgery as post-chemotherapy core
biopsy confirmed the absence of residual cancer. In the remaining
cases, 17 women underwent mastectomy and 17 underwent wide
local excision. The length of time between the final MRI scan and
surgery ranged from 14 to 117 days, with a median of 28 days.
Maximum tumour diameter was recorded in the patient’s notes by
both the pathologist and the MRI radiologist in 21 cases.
Agreement between MRI and pathology measurements
Calculation of the limits of agreement between the maximum
tumour diameters as measured by the pathologists and the MRI
radiologist demonstrated a mean difference (MRI–pathology) of
?2.1mm with a 95% CI (1.96? standard deviation) of 718.6mm.
The limits of agreement were, therefore, ?20.7 to 16.5mm. A
paired t-test of the two sets of measurements showed that the mean
difference was not statistically significant (P¼0.32), thus indicat-
ing that MRI is an unbiased estimator of true tumour size
and, therefore, suitable for quantifying tumour response. The
maximum diameter as measured by the MRI radiologist was also
compared to the cube root of the final tumour volume, and these
twovariablesdemonstrated
(P¼0.0001; R2¼0.56).
excellentlinearcorrelation
Predicting response using individual variables
The ranges of all predictive MRI variables are presented in Table 1,
and the results of correlating the predictive MRI and pathology
variables with PC0F(V) are given in Table 2. Six MRI variables
demonstrated P-values less than 0.07 and the data for these,
divided into SD and PR subsets, are presented in Figure 3. These
six variables along with four histopathology variables for which the
P-values were also less than 0.07 underwent ROC curve analysis,
and the results of this are also given in Table 2. The slight
differences in the patterns of statistical significance between the
correlation and ROC analyses can be attributed to the different
variables quantifying response in the two types of tests (specifically
that PC0F(V) is continuous whereas PR/SD is dichotomous). Four
variables demonstrated an ROC AUC significantly larger than 0.5:
water T2at TP0and changes in %WMRSmeasured at 135ms, water
T2and tumour volume between TP0and TP2. The ROC plots for
these are shown in Figure 4.
All pathology data
9
(5)
10
22
(14)
2(1)
3(2)
All MRI data
at TP0
All MRI data
at TP2
Figure 2
women recruited into this study (with nine cases having to be excluded
completely). Numbers in parentheses indicate the number of cases where
DCIS was present.
A Venn diagram showing the availability of valid data for the 46
Table 1
study, as grouped by clinical time point (TP0being before treatment and
TP2being after the second course of chemotherapy)
The range of values for all predictive MRI variables used in this
Parameter (units)TP0(n¼34)
3.7–240.0
1.4–18.0
0.4–12.4
0.6–3.5
0.8–3.5
0.3–5.9
68–100%
73–100%
47–130
TP2(n¼22)
1.3–95.0
0.6–8.3
0.4–3.9
1.2–3.1
1.3–3.6
0.9–4.3
30–100%
49–100%
24–104
Volume (ml)
Ktrans(per unit time)
Kep(per unit time)
Ve(arbitrary)
MEF (arbitrary)
ADC (mm2/s)
%WMRSat 135ms
%WMRSat 30ms
Water T2(ms)
ADC¼apparent
MRI¼magnetic resonance imaging.
diffusion coefficient;MEF¼maximumenhancementfactor;
Predicting chemotherapy response with quantitative MRI
DJ Manton et al
430
British Journal of Cancer (2006) 94, 427–435
& 2006 Cancer Research UK
Molecular Diagnostics

Page 5

Table 2Statistical significance of non-parametric correlation analyses between the variables indicated and final tumour volume response PC0F(V)
No. of cases
PR+SD
Correlation
P (sense)
ROC curve
P AUC 95% CI
MRI data at TP0
%WMRSat 135ms
Water T2
Tumour volume
Ve
Ktrans
%WMRSat 30ms
Kep
MEF
Water ADC
24+10
24+10
24+10
24+10
24+10
24+10
24+10
24+10
24+10
S: 0.034 (?)
S: 0.055 (?)
S: 0.059 (?)
S: 0.425
S: 0.439
S: 0.579
S: 0.686
S: 0.720
S: 0.808
0.14
0.03
0.20
0.66
0.73
0.64
[1]
0.56 to 0.91
MRI data at TP2
PC02(tumour volume)[2]
D02(water T2)[3]
PC02(%WMRSat 135ms)[4]
D02(Ktrans)
D02(Ve)
D02(Kep)
D02(MEF)
D02(water ADC)
PC02(%WMRSat 30ms)
16+6
16+6
16+6
16+6
16+6
16+6
16+6
16+6
16+6
S: 0.001 (+)
S: 0.006 (+)
S: 0.025 (+)
S: 0.299
S: 0.352
S: 0.405
S: 0.587
S: 0.736
S: 0.788
0.03
0.04
0.02
0.80
0.79
0.83
0.58 to 1.00
0.60 to 0.98
0.60 to 1.00
Other data
Age at TP0
Cancer grade[5]
Presence of DCIS
PRS score[6]
ERS score
Cancer type
26+11
24+11
24+11
24+11
24+11
24+11
S: 0.958
S: 0.003 (?)
M: 0.006 (+)
S: 0.008 (+)
S: 0.061 (+)
K: 0.740
0.08
0.14
0.09
0.17
0.69
0.66
0.68
0.65
0.49 to 0.88
0.49 to 0.88
LRA models (variables included)
MRI only[1,2,4]
MRI and histopathology[1 to 6]
16+6
16+6
0.0032
0.0004
0.92
1.00
0.79 to 1.00
1.00 to 1.00
The sense of the correlations is also shown. Results of ROC curve analyses (with PR/SD as positive/negative results) are also shown for selected variables. Numbers in square
brackets indicate those variables chosen as inputs to LRA modelling. S¼Spearman rank correlation; M¼Mann–Whitney test; K¼Kruskal–Wallis test; (+) a positive correlation
with lower, or more negative values being associated with more negative values of PC0F(V), that is, PR; (?) a negative correlation with higher, or more positive values being
associated with more negative values of PC0F(V), that is, PR. AUC¼area under the curve; ADC¼apparent diffusion coefficient; DCIS¼ductal carcinoma in situ;
ERS¼oestrogen receptor status; LRA¼logistic regression analysis; MEF¼maximum enhancement factor; MRI¼magnetic resonance imaging; PR¼partial response;
PRS¼progesterone receptor status; ROC¼receiver–operator characteristic; SD¼stable disease.
250
225
200
175
150
125
100
100
100
–100
120
140
75
50
50
–50
40
–40
40
40
30
–30
20
–20
20
10
–10
0
0
0
60
–60
60
70
–70
–50
–40
–30
30
–20
20
10
–10
0
–60
–70
–50
–40
–30
30
–20
20
10
–10
0
–60
–70
80
–80
80
90
–90
25
0
Stable disease (SD)
Tumour volume at TP0 (ml)
%WMRS (135 ms) at TP0 (%)
Water T2 at TP0 (ms)
Relative change in tumour volume at TP2 (%)
Relative change in %WMRS (135 ms) TP2 (%)
Change in water T2 at TP2 (ms)
Partial response (PR)
– Medians
Figure 3
less than 0.07. The medians for both SD and PR subsets are also shown.
Dot plots showing the PR and SD data for six MRI variables that when correlated with a final change in tumour volume demonstrated P-values
Predicting chemotherapy response with quantitative MRI
DJ Manton et al
431
British Journal of Cancer (2006) 94, 427–435
& 2006 Cancer Research UK
Molecular Diagnostics