15
Textural Analysis as a Predictor of Breast Tumour Response to Neoadjuvant Chemotherapy P Gibbs, M Lowry, and LW Turnbull Centre for MR Investigations, Hull Royal Infirmary, Anlaby Road, Hull HU3 2JZ

Textural Analysis as a Predictor of Breast Tumour Response to Neoadjuvant Chemotherapy P Gibbs, M Lowry, and LW Turnbull Centre for MR Investigations,

Embed Size (px)

Citation preview

Page 1: Textural Analysis as a Predictor of Breast Tumour Response to Neoadjuvant Chemotherapy P Gibbs, M Lowry, and LW Turnbull Centre for MR Investigations,

Textural Analysis as a Predictor of Breast Tumour Response to Neoadjuvant Chemotherapy

P Gibbs, M Lowry, and LW Turnbull

Centre for MR Investigations, Hull Royal Infirmary, Anlaby Road, Hull HU3 2JZ

Page 2: Textural Analysis as a Predictor of Breast Tumour Response to Neoadjuvant Chemotherapy P Gibbs, M Lowry, and LW Turnbull Centre for MR Investigations,

Introduction

• Preoperative chemotherapy has become a widely accepted treatment for patients with locally advanced breast cancer. Improvements in both relapse-free and overall survival have been shown [1,2].

• Although 50-80% of patients respond to neoadjuvant chemotherapy, a significant percentage of patients show little or no response. Assessment of tumour response is crucial to patient management and is conventionally assessed by clinical examination and X-ray mammography. However, these methods have limitations, particularly in the presence of dense fibroglandular tissue.

• Since a poor response to treatment will prompt a change to non cross-resistant and novel therapeutic agents a more reliable indicator of early response is required.

• MRI has shown some promise in this area. Esserman et al [3] reported that imaging phenotype has potential value as a predictive marker; wherein 77% of patients with well circumscribed masses showed complete or partial response compared to only 20% of patients presenting with diffuse enhancement.

Page 3: Textural Analysis as a Predictor of Breast Tumour Response to Neoadjuvant Chemotherapy P Gibbs, M Lowry, and LW Turnbull Centre for MR Investigations,

• Esserman’s work reinforces the finding that, although radiological assessment of texture is highly subjective, it is known to be a sensitive feature for the determination of pathology [4].

• Texture analysis is an attempt to quantify and emulate this expert eye. Various textural algorithms have been proposed, but the most commonly used method is the spatial grey level dependence matrix technique [5] due to its ability to study the second order statistics of pixels at different angles and spacing.

• The work presented is this poster investigates the efficacy of textural analysis of high resolution post-contrast images in predicting and evaluating breast tumour response to neoadjuvant chemotherapy.

• Data is taken from a prospective trial [6] assessing the role of pharmacokinetic modelling of dynamic contrast enhanced (DCE) MRI, diffusion weighted imaging, and spectroscopic imaging in neoadjuvant chemotherapy.

Page 4: Textural Analysis as a Predictor of Breast Tumour Response to Neoadjuvant Chemotherapy P Gibbs, M Lowry, and LW Turnbull Centre for MR Investigations,

Methods

• We prospectively studied forty-one women (26-75 years old; mean = 50 years) with inoperable primary breast lesions.

• Chemotherapy consisted of six cycles of cyclophosphamide (600 mg/m2) and epirubicin (60 mg/m2) at 21 day intervals, and continuous infusion of 5-fluorouracil (200 mg/m2/day) over 18 weeks.

• Following completion of chemotherapy patients procedured either to wide local excision (19 cases) or mastectomy (19 cases). Three patients were deemed unsuitable for surgery due to extensive metastases and thus underwent needle core biopsy only.

• Histopathological examination post chemotherapy revealed invasive carcinoma not otherwise specified (NOS) in 16 patients, invasive ductal carcinoma in 13 patients, invasive lobular carcinoma in 5 patients, invasive tubular carcinoma in 2 patients, pure DCIS in 1 patient, and no malignant tissue in 4 patients.

Page 5: Textural Analysis as a Predictor of Breast Tumour Response to Neoadjuvant Chemotherapy P Gibbs, M Lowry, and LW Turnbull Centre for MR Investigations,

MR Imaging Protocol:

• All imaging was performed, using a GE Signa Echo-speed 1.5 T scanner, in the prone position with the breasts suspended in a dedicated breast coil.

• After DCE imaging fat suppressed post-contrast data was obtained using a 3D FSPGR sequence (TR/TE 23-28/4.2 ms, flip angle 30°, field of view 20-32 cm, matrix size 512512, slice thickness 3-6.5 mm, 1 average).

• Voxel volumes ranged from 0.46 mm3 to 2.54 mm3.

• MR imaging was performed at 3 time points – prior to commencement of chemotherapy (TP0), after 2 cycles of chemotherapy (TP2), and after completion of chemotherapy but prior to surgery (TPF).

Page 6: Textural Analysis as a Predictor of Breast Tumour Response to Neoadjuvant Chemotherapy P Gibbs, M Lowry, and LW Turnbull Centre for MR Investigations,

Textural Analysis:

• After acquisition ROIs were then drawn, on all appropriate slices, encompassing the lesion as closely as possible.

• The ROI data was then histogram equalised and reduced to 32 grey levels. Histogram equalisation involves replacing each grey level value with a new value in an attempt to ensure the new grey levels are as equiprobable as possible.

Image Intensity

20016012080400

Fre

quen

cy

40

30

20

10

0

Image Intensity

2824201612840

Fre

quen

cy

40

30

20

10

0

Pre (left) and post (right) equalisation histograms from a breast lesion

Page 7: Textural Analysis as a Predictor of Breast Tumour Response to Neoadjuvant Chemotherapy P Gibbs, M Lowry, and LW Turnbull Centre for MR Investigations,

• Co-occurrence matrices, containing the joint probability of adjacent pixels along a given direction having co-occurring values i and j were calculated.

• Four matrices were calculated, for = 0, 45, 90, and 135 degrees, and combined in an averaged co-occurrence matrix since no directional variations in texture were expected.

• Finally, the 14 textural measures defined by Haralick [5] were computed for each lesion (see box for descriptions).

f1 - Angular Second Moment

f2 - Contrast

f3 - Correlation

f4 - Variance

f5 - Inverse Difference Moment

f6 - Sum Average

f7 - Sum Variance

f8 - Sum Entropy

f9 - Entropy

f10 - Difference Variance

f11 - Difference Entropy

f12 Information Measures

f13 of Correlation

f14 - Maximal Correlation

Coefficient

f(x1, y1)=i

f(x2, y2)=j

P (i,j)

Co-occurrence matrix calculation from histogram equalised image (top left)

Calculated textural parameters

Page 8: Textural Analysis as a Predictor of Breast Tumour Response to Neoadjuvant Chemotherapy P Gibbs, M Lowry, and LW Turnbull Centre for MR Investigations,

Results

• A significant reduction in tumour volume was noted over the time course of the chemotherapy regimen.

• To facilitate comparisons the patients were separated into two groups dependent on their response at TP2 – those who showed less than 50% decrease in tumour volume (poor responders) and those who showed greater than 50% decrease in tumour volume (good responders).

• This cut-off point resulted in 20 non-responders and 21 responders.

• Pre-chemotherapy lesion volume shows borderline significance as an indicator of initial lesion response (p=0.06). As might be expected larger lesions showed a greater reduction in absolute tumour volume (Pearson correlation coefficient = 0.966).

TPFTP2TP0

Tum

our S

ize

(cc)

60

50

40

30

20

10

0

Overall reduction in lesion volume

Correlation of initial volume with absolute change in volume

Tumour Volume at TP0 (cc)

300

200

100

50

40

30

20

10

5

4

3

2

1

Red

uctio

n in

Tum

our V

olum

e at

TP

2

300

200

100

504030

20

10

543

2

1

Page 9: Textural Analysis as a Predictor of Breast Tumour Response to Neoadjuvant Chemotherapy P Gibbs, M Lowry, and LW Turnbull Centre for MR Investigations,

TP0

TP2

TPF

Decreasing tumour volume over the complete course of chemotherapy (from 22.3 cm3 to 4.3 cm3)

Page 10: Textural Analysis as a Predictor of Breast Tumour Response to Neoadjuvant Chemotherapy P Gibbs, M Lowry, and LW Turnbull Centre for MR Investigations,

• Significant (p<0.05) or borderline significant (0.05<p<0.09) differences were seen between the two groups for 11 out of the 14 textural parameters calculated, on images obtained at TP0.

• Borderline differences were only seen on 2 parameters on images obtained at TP2.

• This implies that textural parameters can be used to predict initial response. Using a combination of parameters in a logistic regression model revealed a diagnostic accuracy of 0.82±0.07.

Textural Parameter

Good vs poor responders

TP0 TP2

f1 0.066 0.103

f2 0.044 0.406

f3 0.048 0.385

f4 0.712 0.471

f5 0.048 0.291

f6 0.091 0.552

f7 0.088 0.332

f8 0.126 0.134

f9 0.071 0.069

f10 0.052 0.291

f11 0.042 0.228

f12 0.067 0.084

f13 0.080 0.112

f14 0.064 0.223

Comparison of groups using independent-samples t-test (p-values only reported)

1 - Specificity

1.00.75.50.250.00

Sen

sitiv

ity

1.00

.75

.50

.25

0.00ROC curve for predicting

initial response using texture at TP0

Page 11: Textural Analysis as a Predictor of Breast Tumour Response to Neoadjuvant Chemotherapy P Gibbs, M Lowry, and LW Turnbull Centre for MR Investigations,

Changes in Texture for Different Groups:

• More significant changes in texture occur over 2 cycles of chemotherapy for the poor responders compared to the good responders.

• Textural convergence appearance to be occurring.

Box-plot of difference entropy at TP0 and TP2

Good respondersPoor responders

f11

- Diff

eren

ce E

ntro

py

2.8

2.6

2.4

2.2

2.0

1.8

1.6

f11 (TP0)

f11 (TP2)

Textural Parameter

TP0 vs TP2

Poor responders

Good responders

f1 0.347 0.737

f2 0.039 0.937

f3 0.036 0.978

f4 0.245 0.188

f5 0.246 0.298

f6 0.241 0.388

f7 0.029 0.584

f8 0.012 0.042

f9 0.245 0.818

f10 0.036 0.735

f11 0.121 0.686

f12 0.222 0.858

f13 0.089 0.546

f14 0.068 0.796

Comparison of groups using paired-samples t-test (p-values only reported)

Page 12: Textural Analysis as a Predictor of Breast Tumour Response to Neoadjuvant Chemotherapy P Gibbs, M Lowry, and LW Turnbull Centre for MR Investigations,

Clinical Response:

• Clinical response is defined as a decrease in size by 50% or more. Size is taken as the product of the two largest orthogonal diameters and is proportional to volume2/3.

• Using this criteria as a cut-off then 12 patients are defined as non-responders, 27 patients as clinical responders, and 2 patients were excluded since no post-chemotherapy MRI scan was performed.

• 12 out of 14 parameters showed no difference in texture at TPF possibly indicating remaining tissue is chemotherapeutic resistant.

• The smaller lesion volume evident at TPF may account for the poorer quality of the data since counting statistics are reduced. This is especially noticed in the greater range of values for the clinical responders (i.e. those of smaller volume)

Clinical respondersNon-respondersf14

- Inf

orm

atio

n M

easu

re o

f Cor

rela

tion

2

1.0

.9

.8

.7

.6

Box-plot of f14 post-chemotherapy

Page 13: Textural Analysis as a Predictor of Breast Tumour Response to Neoadjuvant Chemotherapy P Gibbs, M Lowry, and LW Turnbull Centre for MR Investigations,

Discussion

• Distinct differences in texture prior to treatment have been noted between lesions that showed a greater initial response, compared to those that had a poorer response.

• Surprisingly, comparing textural parameters at TP0 and TP2 for both good and poor responders revealed more significant changes occurring in the non-responding group. Therefore initial good responders undergo volume changes but do not appear to exhibit textural changes whilst initial poor responders undergo textural changes prior to volume changes.

• The combination of variables in a logistic regression model (as demonstrated herein) or a neural network analysis may aid the determination of patients most suitable for neoadjuvant chemotherapy.

• Co-occurrence matrices are normalised and therefore are ideally independent of the number of pixels present in the lesion. However, smaller lesions lead to reduced counting statistics and thus the textural parameters calculated at TPF can be considered to be less robust than those calculated at TP0.

Page 14: Textural Analysis as a Predictor of Breast Tumour Response to Neoadjuvant Chemotherapy P Gibbs, M Lowry, and LW Turnbull Centre for MR Investigations,

• The number of grey levels defined in the histogram equalisation process is largely a matter of user preference. Reducing the number of grey levels would improve the counting statistics in the co-occurrence matrices, but leads to a concomitant decrease in discriminatory power.

• Further work, especially comparison with histopathological results and pre-chemotherapy/post chemotherapy tumour grade, is necessary to elucidate these results.

• Yorkshire Cancer Research for their continued financial support of the MRI Centre

• Dr David Manton for his input into various aspects of this work

Acknowledgements

Page 15: Textural Analysis as a Predictor of Breast Tumour Response to Neoadjuvant Chemotherapy P Gibbs, M Lowry, and LW Turnbull Centre for MR Investigations,

References

1. Swain A et al. Neoadjuvant chemotherapy in the combined modality approach of locally advanced nonmetastatic breast cancer. Cancer Research 1987;47:3889-94.

2. Hortobagyi GN et al. Management of stage III primary breast cancer with primary chemotherapy, surgery and radiation therapy. Cancer 1988;62:2507-16.

3. Esserman L et al. MRI phenotype is associated with response to doxorubicin and cyclophosphamide neoadjuvant chemotherapy in stage III breast cancer. Annals of Surgical Oncology 2001;8:549-59.

4. Lerski RA et al. MR image texture analysis – an approach to tissue characterisation. Magnetic Resonance Imaging 1993;11:873-87.

5. Haralick RM et al. Textural features for image classification. IEEE Transactions on Systems Man and Cybernetics 1973;3:610-21.

6. Lowry M et al. Neoadjuvant chemotherapy in breast cancer: early prediction of response using a combination of DCE-MRI, ADC mapping and proton spectroscopic imaging.