Motion correction in diffusion-weighted MRI of the breast at 3T

Abstract

Purpose

To provide a quantitative assessment of motion and distortion correction of diffusion-weighted images (DWIs) of the breast and to evaluate the effects of registration on the mean apparent diffusion coefficient (mADC).

Materials and Methods

Eight data sets from 4 patients with breast cancer and 8 data sets from 6 healthy controls were acquired on a 3T scanner. A 3D affine registration was used to align each set of images, and principal component analysis was used to assess the results. Variance in tumor ADC measurements, tumor mADC values, and voxel-wise tumor mADC values were compared before and after registration for each patient.

Results

Image registration significantly (p = 0.008) improved image alignment for both groups and significantly (p < 0.001) reduced the variance across individual tumor ADC measurements. While misalignment led to potential under- and over-estimation of mADC values for individual voxels, average tumor mADC values did not significantly change (p > 0.09) after registration.

Conclusion

3D affine registration improved the alignment of DWIs of the breast and reduced the variance between ADC measurements. Although the reduced variance did not significantly change tumor region-of-interest measures of mADC, it may have a significant impact on voxel-based analyses.

INTRODUCTION

Diffusion-weighted imaging (DWI) provides a means of measuring apparent diffusion coefficient (ADC) values in vivo (1). The ADC describes the rate of random, thermally-induced motion of water within tissue, and its value is dependent on the number of and separation of barriers encountered by the water. Both increased cellularity and reduced extracellular volume (ECV) have been shown to be associated with a reduction in ADC (2). The ADC of water has recently gained increasing attention as a potential imaging biomarker in clinical research of breast cancer (3). One of the hallmarks of cancer is unregulated cell replication (4), which can lead to increased cellularity and reduced ECV in tumors compared to healthy tissue. Preclinical studies have shown that exposure of tumors to both chemotherapy and radiotherapy consistently leads to measurable increases in water diffusion in the case of favorable response (5,6). Preliminary studies in humans have shown that both normal tissue and benign lesions have significantly higher ADC values than malignant breast lesions (7–11) and that successful treatment with neoadjuvant chemotherapy results in an increase in tumor ADC (12,13).

ADC maps are calculated voxel by voxel from at least two images with different diffusion weightings (e.g., b₁ = 50 s/mm² and b₂ = 600 s/mm²) along the same direction. In isotropic media, diffusion-weighted images (DWIs) with gradients applied along only a single direction are required. In anisotropic media, DWIs are acquired with gradients applied along multiple directions (e.g., x, y, and z), and the resulting ADC maps are averaged to produce a rotationally invariant mean ADC (mADC) map. Although diffusion at a given location within a tumor is usually assumed to be relatively isotropic, as there is little direction dependent structure, it is common practice to calculate an mADC map from DWIs with gradients applied along three orthogonal directions to reduce bias from potential directional dependencies (14).

Accurate calculation of the individual and mean ADC maps requires that all these images be co-registered. That is, the same anatomical location is represented by the same image voxel across all the images. Misalignment of the original diffusion-weighted images results in erroneous ADC values because they are calculated with DWI signal values from different locations within the tissue. In practice, distortion caused by eddy currents and bulk patient motion often results in misalignment of these images (15). DWI requires the use of strong magnetic gradients that are switched on and off rapidly, which can induce circulating eddy currents in the conducting material in the MR scanner. These eddy currents produce time- and spatially-varying magnetic fields that interact with the imaging gradients, resulting in shearing, scaling, and translation of the MR signal that vary from image to image. Echo planar imaging (EPI), the most common image acquisition scheme used for DWI, is particularly vulnerable to eddy current distortions in the phase-encode direction because of the relatively low bandwidth per voxel. The most common source of image misalignment is bulk patient motion, which occurs independent of the image acquisition scheme employed. In the brain, motion results in misalignment that can be adequately described by a rigid body transformation. However, in deformable tissue, such as the breast, severe motion may result in more complex deformations.

A large body of work addressing these issues exists for the brain. Methods designed to prevent (16) or reduce (17) the effects of eddy currents have been proposed, and several image post-processing methods have been developed to remove them or reduce the effects of residual eddy current distortion (18–20). Haselgrove and Moore (19) were the first to propose alignment of the diffusion-weighted images to a non-diffusion-weighted image volume as a means of correcting distortions induced by eddy-currents. Alternative image registration methods make use of the fact that the distortion can be described by an affine transformation on a slice-by-slice basis (20). Andersson and Skare (18) first proposed the use of image registration to simultaneously correct for eddy current distortions as well as patient motion in DWIs of the brain.

Although motion and distortion corrections have now become common practice in DWI of the brain, little attention has been paid to these problems in DWI of the breast until recently (21–23). Partridge et al. (22) commented that alignment of DWIs of the breast with a nonlinear registration algorithm improved the quality of their data, but the authors provided no quantitative measure of the improvement. To our knowledge, no studies have been published that provide a quantitative measure of the effects of image alignment on the quality of ADC maps derived from DWIs of the breast. However, the effects of image misalignment on the accuracy of ADC estimation have important consequences in quantitative studies of ADC in tumors. While the diffusion at a given location within a tumor may be relatively isotropic with similar ADC values measured along each of the gradient directions, diffusion properties may be heterogeneous throughout a tumor, depending on its state. For example, regions of necrosis may have significantly higher ADC compared to regions of high cell density (9). If the individual DWIs are not well aligned and the tumor exhibits heterogeneous diffusion properties, inaccurate ADC maps will result.

Quantifying the accuracy of image registration is a challenging task, because a gold standard is often not available. One way to address this problem is to measure the overall misalignment by manually locating several landmarks in the target and registered images and measuring the root-mean-squared error between them. This method is time-consuming and vulnerable to user-error; however, it is relatively straightforward to perform in the brain, where there are several easily defined anatomical landmarks. In other soft tissue structures, such as the breasts, this method is far more challenging to implement. A potential alternative is the use of principal component analysis (PCA). PCA is a fully automated analysis method that has been shown to provide a quantitative means of evaluating DWI registration results in the brain (15,24). Rohde et al. (15) observed that image distortion and subject motion induced changes in the percentage of total variance in the first four principal components (PCs) of DWIs and that these values changed significantly in a predictable manner after image alignment. Maniega et al. (24) used percentage of total variance described by each PC to quantify differences in alignment of DWIs using two different image registration algorithms. Because PCA does not require the definition of anatomic landmarks, it can be used to evaluate the registration of any type of image, including breast DWIs, through changes in the values of the percentage of total variance explained by the individual PCs before and after registration.

Although bulk subject motion during DWI of the breast may result in nonlinear deformation, we hypothesized that affine registration, commonly used for DWI of the brain, would improve the alignment of DWIs of the breast. To test this hypothesis, we performed a PCA and compared the percent of total variance explained by the individual principal components of the registered and unregistered DWIs from eight patient data sets and eight control data sets. We also hypothesized that improvements in image alignment would reduce the voxel-wise variance in tumor ADC measurements across diffusion directions, affecting statistical analyses commonly performed on these data sets, such as region of interest (ROI) and voxel-wise analysis. We tested this hypothesis by first comparing the voxel-wise variance of the tumor ADC values measured in three different directions before and after registration. In addition, mADC values averaged over the tumor for each patient were compared before and after registration, and mADC residuals were tested on a voxel-wise basis for identification of potential outliers.

MATERIALS AND METHODS

Subjects

Data were acquired from four patients with invasive ductal carcinoma (mean age = 45 years, range = 39 to 50 years) as part of an ongoing IRB approved clinical trial. Each patient participated in two separate scanning sessions, resulting in a total of eight patient data sets. The timing of the two scan sessions depended on the treatment plan of the patient. Two patients were scanned prior to neoadjuvant chemotherapy and again after completion of the first cycle of treatment (Patients 1 and 4). The other two patients were scanned on consecutive days prior to mastectomy (Patients 2 and 3). Six control subjects (mean age = 49 years, range = 34 to 61 years) participated in a single scanning session. The DWI sequence was repeated for two of the control subjects, resulting in a total of eight control data sets.

Image Acquisition

Diffusion weighted images were acquired with a 3T Achieva MR scanner (Philips Healthcare, Best, The Netherlands), using a double-breast 4-channel sensitivity encoding (SENSE) receive coil (Invivo Inc., Gainesville, FL) and 80 mT/m gradients (100mT/m/ms slew-rate). DWIs were acquired with a single-shot spin echo (SE) echo planar imaging (EPI) sequence in three orthogonal diffusion encoding directions (x, y, and z), with two b-values (0 and 600 s/mm²), FOV = 192×192 mm², acquisition matrix of 96×96, reconstruction matrix of 144×144, and 10 averages. SENSE parallel imaging (acceleration factor = 2) and spectrally-selective adiabatic inversion recovery (SPAIR) fat saturation were implemented to reduce image artifacts. Subjects were breathing freely with no gating applied. The patient DWIs consisted of 12 sagittal slices with slice thickness = 5 mm (no slice gap) and TR = 2255 ms, TE = shortest (43, 48, or 51 ms), diffusion time (Δ) = 20.7, 23.2, or 24.9 ms, and diffusion gradient duration (δ) = 11.6, 10.6, or 10.2 ms, respectively, with a total scan time of 2 min and 42 s. The control DWIs consisted of 16 sagittal slices with slice thickness = 2.5 mm (no slice gap) and TR = 3000 ms, TE = shortest (52 ms), Δ = 25.2 ms, δ = 10.1 ms, and a total scan time of 3 min and 36 s.

Image Registration

Motion and eddy current distortion correction were performed using the Diffusion Registration tool available on the scanner console (Release 2.6.3.4). The tool registers each diffusion-weighted (b > 0 s/mm²) image volume to the corresponding non-diffusion-weighted (b = 0 s/mm²) image volume using a 3D affine transformation with a local correlation (LC) similarity measure optimized by the Gauss-Newton method. A brief description of the process and the parameter settings are described below. Details of the similarity measure and optimization scheme are described by Netsch and van Muiswinkle (25).

The LC value is calculated by defining a neighborhood surrounding each voxel in the reference image (b = 0 s/mm²), calculating the correlation coefficient between the reference neighborhood and the corresponding neighborhood in the image being transformed (b > 0 s/mm²), and summing these values over the image. It has been shown that only voxels with high intensity variance within their local neighborhood contribute significantly to the LC calculation and use of this fraction of the total image voxels in the optimization process reduces processing time and produces a reliable result (26). The number of voxels selected for LC calculation with respect to the total number of image voxels is referred to as the LC fraction, and typically ranges from 5–20%. The iterative optimization process is repeated until either a maximum number of steps has been performed or until the mean displacement of the outermost eight voxels considered in the registration process is less than a given minimum distance for an updated optimization step. The Diffusion Registration tool performs the LC calculation with voxel neighborhoods of 3×3×3 and an LC fraction of 10%, and the optimization process is performed with a maximum of 20 steps and a minimum distance of 0.5 mm. Linear interpolation is used for image resampling in both the optimization process and the final image transformation.

Quantitative Assessment of Image Registration Results

The results of a principal component analysis (27) were used to compare the alignment of the registered and unregistered DWIs for each subject (patients and controls). First, a single slice encompassing the largest cross-section of tumor, in the case of the patients, or glandular tissue, in the case of the controls, was selected. Then, a mask was manually defined to exclude signal from the background and the chest cavity, so that only voxels within the breast and the chest wall were included in the PCA. This masking step was performed to reduce PCA artifacts potentially introduced by severe motion-induced image artifacts within the chest cavity from the heart and lungs. Finally, for each image set from each subject, the PCA was performed using a total of four images: the non-diffusion-weighted image (b = 0 s/mm²) and the three diffusion-weighted images (b = 600 s/mm²) applied in the x, y, and z directions.

For each of the four images, i, the intensity values from the N voxels within the mask were arranged into a vector, x_i, where x_i = [x₁,x₂,..,x_N]^T. To avoid bias from the difference in maximum intensity values between the b = 0 s/mm² and the b = 600 s/mm² images, the elements of x_i were then standardized by subtracting the mean and then dividing by the standard deviation. The standardized vectors were then combined to form an N × 4 array:X=[x₁,x₂,x₃,x₄]^T. The principal components (PCs) of the image set were then calculated by computing the eigenvalues, λ, and eigenvectors of the correlation matrix of X (24). Each eigenvalue and its associated eigenvector represent a single PC. The PCs are typically ranked in descending order, where the first PC corresponds to the largest eigenvalue.

The eigenvalues are a quantitative measure of how representative the corresponding PC is of the data. For DWIs, the first PC is primarily related to the T₂-weighted contrast in the images, the second PC is related to the contrast due to isotropic diffusion, and the higher PCs are related to noise and contrast due to anisotropic diffusion (15,24). The sum of the eigenvalues is a measure of the total variance in the image set, and the percent of total variance (%Var) explained by each PC is calculated as follows:

$$\%{\text{Var}}_i=\frac{{\lambda }_i}{{\displaystyle {\sum }_{j=1}^4{\lambda }_j}}.$$ [1]

The %Var values were compared between the unregistered and registered image sets for each subject using a Wilcoxon rank test because the sample statistics for the %Var values was unknown.

Effects of Registration on the Tumor Mean Apparent Diffusion Coefficient

To investigate the hypothesis that image registration reduces the variance in individual ADC measurements, the variance in tumor ADC values measured in the three diffusion-weighting directions (x, y, and z) before and after registration was compared for each of the patient data sets. First, slice-matched, post-contrast T₁-weighted images were used in manual definition of tumor regions of interest (ROIs) for each subject within the slice used in the PCA. The sample variance for each voxel within the ROI was then computed according to the following equation:

$$s^2=\frac{1}{N-1}{\displaystyle \sum _{k=x,y,z}{({\text{ADC}}_k-\text{mADC})}^2},$$ [2]

where

$$\text{mADC}=\frac{{\text{ADC}}_x+{\text{ADC}}_y+{\text{ADC}}_z}{3}$$ [3]

and N=3. The pre- and post-registration variances were compared using a Wilcoxon signed rank test, as the sample statistics were unknown.

Next, the effects of changes in ADC measurement variance on average tumor mADC values and voxel-wise mADC values were investigated. For each patient, mADC values were averaged over the tumor ROI and the pre- and post-registration values were compared using a Student’s t-test. In addition, 95% confidence intervals (CIs) for the mADC residuals (the difference between mADC values pre- and post-registration) were calculated within the tumor ROI for each patient, and voxels exhibiting a residual value outside the CIs were identified.

RESULTS

Quantitative Assessment of Image Registration Results

For both the patient group and the control group, the percentage of total variance explained by each PC changed significantly (p = 0.008) after image registration. For all subjects, the %Var increased in the first PC and decreased in the higher PCs. Values of %Var before and after registration are plotted in Figure 1. The corresponding eigenvectors were reformatted to produce eigenimages for qualitative comparisons, as well. A representative example is shown in Figure 2. While the pre- and post-registration %Var values were significantly different for both the first and fourth PCs, the differences between their pre- and post-registration eigenimages are quite subtle. The most notable effects of motion were seen in the pre-registration eigenimages of the second and third PCs. Contrast in the eigenimage of the second PC should be due primarily to isotropic diffusion; however, contrast at the tissue borders, such as the breast/air and breast/muscle borders (white arrows), was introduced by motion. Contrast in the remaining eigenimages should be due primarily to noise; however, anatomical features, such as the tumor (double arrowheads) and glandular tissue, were introduced by motion in these eigenimages. These artifacts were removed from the eigenimages after image registration.

Figure 1.

Plots of the percentage of total variance explained by each of the first four PCs before (Pre) and after (Post) image registration for the controls (a) and patients (b). For both groups, the percentage of total variance increased in the first PC and decreased for the higher PCs.

Figure 2.

PC images for the 1st (a–b), 2nd (c–d), 3rd (e–f), and 4th (g–h) PCs from Patient 1 (Day 0). PC images resulting from the unregistered and registered DWIs are shown in the left and right columns, respectively. The most notable differences between the unregistered and registered results are in the 2nd and 3rd PCs, which are related to the contrast caused by isotropic diffusion and noise, respectively. Motion in the unregistered images led to artifacts around the border of the breast (closed arrowheads) and the tumor (double arrowhead). The corresponding post-contrast, T₁-weighted anatomical image is shown for reference (i).

Effects of Registration on the Tumor Mean Apparent Diffusion Coefficient

The results of the comparisons of pre- and post-registration variance and average ROI mADC values are shown in Table 1. The variance in the tumor ADC values significantly decreased after image registration (p < 0.001) for all patients. However, in general, mADC values averaged over the tumor ROI did not significantly change after image registration. The mADC value increased an average of 1.5% in six of the patient data sets, and decreased by 2% and 8% in the remaining two data sets. The 8% decrease in mADC in Patient 3 (Day 7) approached significance (p = 0.02), and the changes measured in the remaining data sets had p-values ranging from 0.09 to 0.9.

Table 1.

Results from comparison of ROI statistics before and after registration of individual ADC maps

		var(ADC) (µm4/ms2)			mADC (µm2/ms)
Patient (Day)	ROI size (voxels)	Before	After	p	Before	After	p
1 (0)	535	3.17 × 10−2	1.02 × 10−2	3.79 × 10−54	1.37	1.35	0.23
1 (14)	404	1.40 × 10−2	9.10 × 10−3	1.13 × 10−09	1.30	1.31	0.30
2 (0)	126	5.99 × 10−3	3.81 × 10−3	8.25 × 10−04	1.05	1.06	0.78
2 (1)	190	6.42 × 10−3	4.53 × 10−3	5.47 × 10−06	1.13	1.15	0.62
3 (0)	189	1.20 × 10−2	6.80 × 10−3	1.18 × 10−03	1.07	1.08	0.76
3 (7)	180	8.36 × 10−3	2.49 × 10−3	1.76 × 10−14	1.03	0.95	0.02
4 (0)	433	1.42 × 10−2	8.13 × 10−3	6.51 × 10−25	1.37	1.37	0.91
4 (7)	194	1.26 × 10−2	5.78 × 10−3	4.10 × 10−12	1.16	1.13	0.45

Significant differences are denoted by bold text. var(ADCi): variance of ADC_x, ADC_y, and ADC_z values averaged over the ROI. mADC: mean ADC value average over the ROI.

The 95% CI for the mADC residuals and the percentage of ROI voxels with residuals outside the CIs are shown for each subject in Table 2. CI values ranged from ±0.11 µm²/ms to ±0.60 µm²/ms. The percent of ROI voxels with residuals outside their respective CIs ranged from 9–13%, and those voxels were found to lie mostly along the ROI borders.

Table 2.

95% confidence intervals for mADC residuals and the percentage of ROI voxels with residuals outside the confidence intervals

	95% CI for residuals (µm2/ms)	% of ROI voxels outside 95% CI
Patient 1, Day 0	±0.37	12%
Patient 1, Day 14	±0.18	13%
Patient 2, Day 0	±0.16	12%
Patient 2, Day 1	±0.18	12%
Patient 3, Day 0	±0.40	13%
Patient 3, Day 7	±0.60	10%
Patient 4, Day 0	±0.11	6%
Patient 4, Day 7	±0.34	9%

CI: Confidence interval

The results for a single data set (Patient 1, Day 0) are shown in Figures 3 and 4. Pre- and post-registration maps of tumor ADC variance and mADC are displayed in Figure 3. Before registration, the variance in tumor ADC measurements was quite high (≥0.1 µm⁴/ms²) along the anterior margin of the tumor (Panel B). After registration, the variance was greatly reduced throughout the tumor (Panel C). A few voxels along the tumor margin exhibited extreme mean ADC values (<0.5 µm²/ms or >2 µm²/ms) prior to registration (Panel D); however, image alignment did not result in an overall change in the tumor mADC value (Panel E). A scatter plot of the pre- and post-registration mADC values for the tumor ROI voxels is shown in Figure 4 (left panel). The 95% CI of the residual values was ±0.37 µm²/ms (dashed lines), and outliers (voxels with residuals less than −0.37 µm²/ms or greater than 0.37 µm²/ms) are plotted in red. The corresponding voxel locations are mapped within the tumor ROI and found to lie primarily along the border of the ROI (center panel).

Figure 3.

Comparison of ROI statistics for Patient 1 (Day 0). (a) The tumor ROI (solid white line) was manually defined in the slice closest to the center of the tumor in the non-diffusion-weighted image (b = 0 s/mm²). Maps of the variance across individual ADC measurements (ADC_x, ADC_y, and ADC_z) for each voxel in the ROI are shown for the unregistered (b) and registered (c) DWIs, respectively. Maps of the mean ADC values for each voxel in the ROI are shown for the unregistered (d) and registered (e) DWIs, respectively.

Figure 4.

95% confidence intervals for the mADC residuals of Patient 1 (Day 0). a: The pre-registration values of mADC values for each voxel are plotted against the post-registration values. The 95% CIs for the residual mADC values after registration are plotted for reference (dashed lines). Voxels with a residual value outside the 95% CIs are plotted in red. b: The voxel location corresponding to the residual outliers are marked in red and lie primarily along the border of the tumor ROI. c: The b = 0 s/mm² image with the tumor ROI is shown for reference.

DISCUSSION

We have presented a quantitative analysis of the effects of affine registration of DWIs of the breast. Partridge et al. (22) commented that although the subjects in their studies typically remain relatively still, image registration resulted in a qualitative improvement in the quality of their DWI data. As interest in using ADC as a quantitative biomarker for tumor detection and treatment assessment increases, it is increasingly important to evaluate data quality and the quality of the methods being used in those studies. This is especially true at higher fields, where an increase in magnetic susceptibility effects can introduce additional nonlinear image distortion.

In this study, eddy current distortion and bulk subject motion were difficult to detect by visual inspection alone. Motion appeared to primarily be due to chest wall movement during the cardiac cycle and respiration. Active shielding and eddy current compensation were implemented on the scanner used in this study, reducing the effects of eddy currents, and the lack of motion exhibited by the subjects may have been a result of participant selection.

In spite of a lack of visible gross bulk motion and eddy current distortion, significant changes in the %Var for each PC were found in the registered images compared to the unregistered images for both the patient and control groups. Misalignment of the unregistered images resulted in an artificial decrease in the %Var of the first PC and increase in the %Var values of the remaining PCs. In other words, the misalignment introduced artifacts in the PCA reducing the amount of variance explained by the T₂-weighted contrast in the images and increasing the amount of variance explained by diffusion and noise. Affine registration of the images resulted in a significant increase in the %Var for the first PC and a significant decrease in the %Var for the higher PCs, indicating that the registration had improved the alignment of the images, as we hypothesized.

PCA does not provide an absolute measurement of the accuracy of the registration results because the %Var values for the PCs depend upon the scan parameters. For example, the %Var values for the first PC of controls (top row) are generally lower than those of the patients (bottom row), and the %Var values for the higher PCs are generally higher for the controls than the patients, as seen in Figure 1. This is most likely due to the difference in slice thickness between the two groups. Although the control and patient DWIs had the same in-plane voxel size (1.33 mm × 1.33 mm), the control DWIs had a smaller slice thickness (2.5 mm) than the patients (5 mm). This resulted in a smaller volume per voxel, and thus lower signal to noise, for the controls than the patients. Therefore, more of the variance is explained by noise (higher PCs) in the control data than the patient data. Despite this limitation, PCA is a useful tool for comparing pre- and post-registration results or registration results from different algorithms (15,24).

Improved alignment of the images resulted in a significant decrease in the variance between tumor ADC values measured along different directions. This reduced variance should theoretically result in a more accurate estimation of mADC in tissues that exhibit relatively isotropic diffusion. In this study, motion and eddy current induced distortions were not severe enough to result in significant differences in the mean of the tumor ROI mADC values before and after registration. This suggests that mild tissue deformation induced by cardiac and respiratory motion is not enough to significantly affect the results of a ROI analysis of mADC. However, it should be noted that a large number of signals were acquired and averaged during image reconstruction in this study to improve the signal-to-noise ratio, potentially averaging out the effects of motion and distortion. Takahara et al. have demonstrated excellent visualization of tumors using the diffusion weighted whole body imaging with background body signal suppression (DWIBS) technique with free breathing (28–30), where a large number of signals are acquired and averaged to improve the quality of the data and reduce the effects of motion. However, the DWIBS technique is primarily used as a visualization tool (29). Further study is needed to investigate the effects of subject motion between acquisition of images averaged during image reconstruction on the accuracy of ADC estimation.

There are several clinical applications for quantitative measurements of tumor mADC in breast cancer, including tumor differentiation (7,9,31) and assessment of treatment response (12,13). The most common analysis method used to compare mADC values across time or across different regions of tissue is ROI analysis. Typically, in an ROI analysis, statistical tests are performed to investigate differences between the average mADC value across ROIs. By examining changes in the ROI average value, and not changes in the value of the individual voxels, this effectively smoothes the data and removes the effects of small changes in variance. While ROI analyses are straightforward and relatively simple to implement, they are not without limitations. For example, partial volume averaging in heterogeneous tissues may average out significant changes, and inconsistent ROI definition may introduce error. It is possible that drastic misalignment of the DWIs is necessary before significant error is introduced in a ROI analysis. However, voxel-wise analyses, such as functional diffusion mapping (fDM) (32), may be more appropriate in longitudinal studies to predict outcome and assess treatment response, and they may be more sensitive to the effects of image misalignment. In this study, it was found that mADC residuals (post-registration mADC – pre-registration mADC) for approximately 10% of the voxels within a given tumor ROI lie outside the respective 95% confidence interval. This suggests that the mADC values for these voxels may have been either over- or under-estimated prior to image alignment. However, this is difficult to verify, as we do not know the true ADC values for the tumors.

One of the limitations of this study is the small sample size: 8 data sets acquired from 4 patients before and after the start of treatment and 8 data sets acquired from 6 healthy controls where 2 controls repeated the DWI scan within the same imaging session. There is a risk that multiple measurements from the same subjects could potentially be highly correlated. However, patients usually experience increased discomfort during scan sessions after the start of neoadjuvant treatment, so we typically expect more patient motion in the post-treatment scans than the pre-treatment scans. Also, inclusion of the two same-scan session control data sets did not affect the PCA results. Removing those two scans from the analysis still resulted in a significant p-value (p=0.03). We also assumed relatively isotropic diffusion within the breast tissue. If the diffusion is truly isotropic, then an acquisition scheme where the diffusion-weighting gradients are applied simultaneously (33,34) could provide a much shorter scan time thereby reducing patient motion. If the diffusion is anisotropic, then better spatial resolution and more than 20 unique diffusion-encoding directions would be necessary to accurately characterize diffusion anisotropy (35). Partridge et al. (23) recently investigated diffusion anisotropy in the breast; however, they reported only low to moderate anisotropy that varied with location within the breast, and they only acquired 6 diffusion directions.

In conclusion, we have demonstrated that affine image registration improves the alignment of individual DWIs of the breast. While ROI analysis of tumor mADC is relatively immune to minor misalignment of maps of ADC values measured along different directions, we recommend the use of image registration to align DWIs of the breast prior to ADC quantification, particularly for studies implementing voxel-based analysis methods.