Addition of T2-guided optical tomography improves noncontrast breast magnetic resonance imaging diagnosis

The imaging protocol for human subject participation was approved by the Committee for the Protection of Human Subjects at Dartmouth and Xijing Hospital in Xi’an, China. Women presenting with a clinical breast abnormality scheduled for surgical resection without known contraindications for MRI or Gd injection were invited to participate. Written informed consent was obtained from each participant. Twenty-four women were involved in the study and fell into an age range from 24 to 64 years. Of these 24 patients, 16 were found to have pathologically confirmed breast cancer and 8 had benign conditions. The average ages of the malignant and benign groups were 48 ± 11 years (range 24–64 years) and 31 ± 8 years (range 20–44 years), respectively. Mean body mass index (BMI) was 23.3 ± 3.6 kg/m² while breast sizes were distributed as 10 A-cup, 8 B-cup, 3 C-cup, and 3 D-cup. Because most women in China do not participate in breast screening programs, patients did not have prior mammography. As a result, breast density was assessed based on MRI, and categorized as 1 fatty, 10 scattered, 7 heterogeneously dense, and 6 extremely dense. Subject data are summarized in Additional file 1 (Table S1).

NIRST was performed with a hybrid frequency domain (FD)-continuous wave (CW) optical imaging system described in detail previously [30, 31]. The device included six FD wavelengths (spanning from 660 nm to 850 nm) and three CW wavelengths (900 nm to 950 nm). NIRST data were acquired at all nine wavelengths simultaneously with MRI by attaching an optical fiber interface to the commercial breast coil system already in place for clinical use [20]. To acquire the optical data, sixteen sequential source positions illuminated the breast through a custom optical switch. During each individual source illumination, the remaining 15 fibers detected transmitted light, yielding a total of 240 measurements at each wavelength. For FD measurement, the amplitude and phase of the detected light were collected by lock-in detection. For CW measurement, only amplitude data were recorded. The total NIRST imaging time per breast was 15 min.

MRI was performed using a Siemens MAGNETOM Trio 3.0 T scanner and involved standard clinical sequences. Specifically, acquisitions included a bilateral T1 precontrast scan (slice thickness < 3 mm), a unilateral T2 turbo spin echo sequence with fat suppression (slice thickness < 5 mm, TR/TE = 6490/61 ms, flip angle = 120°, voxel size = 1.0 × 1.0 × 4.0 mm, FOV = 332 mm², and 100% FOV phase), bilateral DWI with eight b-values ranging from 0 to 1400 s/mm², and a series of five bilateral T1 postcontrast scans spaced 90 s apart. Apparent diffusion coefficient (ADC) maps were calculated automatically by software on the MRI system. To achieve coregistration between MRI and NIRST data, fiducial markers were placed at the end of each NIRST fiber bundle.

MRI examinations were evaluated independently by three radiologists experienced in diagnostic breast MRI (JX, more than 15 years; XL, more than 10 years; and KW, more than 8 years) who were blinded to pathological results; they scored the image data according to the Breast Imaging Reporting and Data System (BIRADS). Noncontrast images were assessed based on morphological features in the T2 scans with ADC values as references. In cases of disagreement, the final MR diagnosis (benign versus malignant) was based on majority consensus between the three radiologists. All lesions were measured along three orthogonal axes, and the greatest diameter was considered in the statistical analysis. In cases of multifocal or multicentric abnormalities, only the largest lesion was evaluated. DCE images were assessed according to Teifke criteria for contrast enhancement in focal breast lesions [32], which incorporates shape, border characteristics, enhancement kinetics, and enhancement pattern and maps to a BIRADS category based on cumulative scores [33]. Suspicious regions were manually segmented by a radiologist (JX) to create regions-of-interest (ROIs) for subsequent optical property assessment using either the T2 and DWI images (for noncontrast MRI analyses) or the DCE results (for comparative DCE MRI analyses). For DCE images, ROI segmentation was based on subtraction images formed by subtracting precontrast images from postcontrast images acquired 78 s after contrast injection. OsiriX image processing software (OsiriX MD 7.0, Pixmeo SARL, Bernex, Switzerland) was used to process all MRI data.

Breast images were processed and reconstructed based on the open-source software platform NIRFAST [34]. Prior to NIRST reconstruction, a patient-specific finite element mesh was generated from T1 MR images (Fig. 1a). Then, data calibration was performed with a reference phantom to correct for small variations in detector response and light delivery, and to obtain initial estimates of optical properties for NIRST image reconstruction. Finally, NIRST image reconstruction was constrained by MR-derived spatial priors encoded through an automated direct regularization method that does not require MR segmentation [35], in which a weighted matrix, L, has the form:

where γ _i and γ _j are the grayscale values in the 16-bit MR image data (Fig. 1b for DCE-guided or Fig. 1c for T2-guided) mapped to nodes i and j on the FEM mesh, r _i and r _j are the coordinate positions of nodes i and j. M _i is a factor chosen for ith row in L, and satisfies \( {M}_i=\sum_{j=1,j\ne i}^n{L}_{ij}\forall i=1,\dots, n \) where n is the number of finite element nodes. max(|r _i  − r _j |) is the maximum distance between any two finite element nodes. The function, θ(⋅), is the Heaviside step function, which determines the local weight applied to the ith NIRST image reconstruction position. σ _g is the characteristic grayscale difference over which to apply regularization, and σ _d is a factor related to the distance of influence of elements in the weight matrix relative to NIRST image position i. The operator, L, encodes optical property uniformity by penalizing similarly gray MR locations in the NIRST image to have similar update values at each iteration of the NIRST image reconstruction algorithm. In this study, parameters in L were fixed based on previous tests [35, 36], and set to be σ _g  = 0.01, and σ _d  = 0.4. The regularization parameters used in our experiments were 10 * max(diag(J _k ^T J _k )), where J _k is the Jacobian matrix at the kth iteration in the update equation:where Δx _k is the update to the chromophore concentrations, d is the measured data, f(x _{k − 1}) is the forward solution using the estimated parameters from the k − 1th iteration, and the superscript T denotes the transpose operation.

After NIRST image reconstruction, total hemoglobin (HbT) contrast was computed as the ratio of the average HbT in the abnormal ROI to the average HbT in the rest of the breast, and used to assess differences in benign and malignant abnormalities in subsequent diagnostic performance analyses. For evaluations based on noncontrast MRI data, ROIs segmented from the noncontrast MRI images were used whereas ROIs segmented from the DCE data were applied in the comparative DCE analysis.

Student’s t tests were performed to access statistical differences in mean values (HbT contrast ratio, BIRADs score, and their combination) for breast abnormalities pathologically classified as benign or malignant. Receiver-operating characteristic (ROC) analysis was completed to evaluate differences in differentiation of benign versus malignant lesions as a function of cutoff value (HbT contrast ratio for NIRST, BIRADS score for MRI). The threshold corresponding to the largest summation of the average sensitivity and specificity on the ROC curve was considered to be the best cutoff point. To assess the combination of MRI and NIRST, multiple logistic regression coefficients were assembled into a single score which minimized the difference between the combined variables and the pathological diagnosis. Significance for all statistical tests was assumed at a confidence interval of 95% (P < 0.05) for a two-tailed distribution. The corresponding sensitivity and specificity in the ROC analysis are reported and NIRST reconstruction results guided by DCE MRI are also listed for comparison.

A woman with an undiagnosed 11 × 21 × 14 mm³ lesion in her right breast, later pathologically confirmed to be malignant, was imaged with MR-guided NIRST. Figure 1 shows NIRST HbT images overlaid on the corresponding T1 scans, based on DCE and T2 guidance. The DCE MRI result in Fig. 1b was formed by subtracting precontrast images from postcontrast images acquired 78 s after contrast injection. The sequence used to acquire the DCE MRI was nonfat-suppressed as shown in Fig. 1a. HbT contrast in the ROI (determined independently from DCE or noncontrast MRI image data) was 1.5 and 2.8, respectively, which is indicative of malignancy.

A woman with a 21 × 34 × 26 mm³ abnormality in her left breast, pathologically confirmed as cystic hyperplasia, underwent a MR-guided NIRST examination prior to definitive diagnosis. Figure 2 shows representative MR and HbT images based on DCE- and T2-guided methods. As in Fig. 1, the DCE MRI acquisition was nonfat-suppressed, and postcontrast subtraction images are presented. HbT contrast in the ROI (determined independently from DCE or noncontrast MRI image data) was 0.9 and 0.5, respectively, for the two guidance methods, which is indicative of a benign lesion.

Figure 3 presents two cases where noncontrast MRI and/or DCE MRI images resulted in misdiagnoses. The top row contains a false negative case in which the subject had a 10 × 25 × 25 mm³ lesion in her right breast that was interpreted as benign based on T2 + DWI MRI. Data from NIRST indicated the abnormality was malignant given its HbT contrast (average HbT in abnormal ROI to the average HbT in the rest of the breast) of 3.3 (higher than the cutoff value of 1.1) extracted from the T2-guided NIRST images. Postsurgical pathological results confirmed the tissue was malignant.

The bottom row of Fig. 3 shows a noncontrast MRI and DCE MRI false positive case where the diagnosis based on either noncontrast MRI or DCE MRI was malignant. In comparison, the T2 MRI did not exhibit much regional enhancement, and was used to guide HbT contrast from NIRST to reveal a value of 0.9 (lower than the cutoff value of 1.1) suggesting a benign diagnosis. Postsurgical pathological analysis confirmed the tissue was benign (adenosis).

Diagnostic performances of MRI alone, MR-guided NIRST alone, and the combination of MRI with NIRST are summarized in Table 1. Assuming a BIRADS 4 (or higher) image rating to be positive for cancer, noncontrast-enhanced MRI (T2 + DWI) yielded sensitivity, specificity, and diagnostic accuracy of 88%, 88%, and 88%, respectively, compared to values of 100%, 63%, and 88% for DCE MRI using the same cut-off (i.e., BIRADS ≥ 4 as a cancer diagnosis). ROC analysis of MRI alone yielded area under the curve (AUC) of 0.88 and 0.81 in the noncontrast MRI and DCE MRI cases. Standard deviations in the sensitivity, specificity, and accuracy results from the MRI assessments performed independently by the three radiologists were 6%, 0%, and 4%, respectively.

For T2- and DCE-guided NIRST, a significant difference was found in mean HbT contrast (P ≤ 0.001) between pathologically confirmed malignant and benign abnormalities in the study participants (Fig. 4). For an HbT contrast cutoff value of 1.1, sensitivity, specificity, and diagnostic accuracy were 88% for both T2- and DCE-guided NIRST breast examinations. ROC analysis (Fig. 5) yielded AUCs of 0.91 and 0.90 in the two guidance cases.

When noncontrast MRI (T2 + DWI) and T2-guided NIRST were combined, sensitivity, specificity, diagnostic accuracy, and AUC increased to 94%, 100%, 96%, and 0.95, respectively. When DCE MRI was combined with DCE-guided NIRST, sensitivity, specificity, and accuracy remained at 88%, although AUC increased to 0.94.

No statistically significant difference was found in the diagnostic performance of noncontrast MRI relative to DCE MRI (P = 0.6), or from the combination of noncontrast-enhanced MRI (T2 + DWI) and T2-guided NIRST relative to the combination of DCE MRI and DCE-guided NIRST (P = 0.89).