Automated analysis of the total choline resonance peak in breast proton magnetic resonance spectroscopy

The aim of the current study was to compare the performance of fully automated software with human expert interpretation of single‐voxel proton magnetic resonance spectroscopy (1H‐MRS) spectra in the assessment of breast lesions. Breast magnetic resonance imaging (MRI) (including contrast‐enhanced T1‐weighted, T2‐weighted, and diffusion‐weighted imaging) and 1H‐MRS images of 74 consecutive patients were acquired on a 3‐T positron emission tomography‐MRI scanner then automatically imported into and analyzed by SpecTec‐ULR 1.1 software (LifeTec Solutions GmbH). All ensuing 117 spectra were additionally independently analyzed and interpreted by two blinded radiologists. Histopathology of at least 24 months of imaging follow‐up served as the reference standard. Nonparametric Spearman's correlation coefficients for all measured parameters (signal‐to‐noise ratio [SNR] and integral of total choline [tCho]), Passing and Bablok regression, and receiver operating characteristic analysis, were calculated to assess test diagnostic performance, as well as to compare automated with manual reading. Based on 117 spectra of 74 patients, the area under the curve for tCho SNR and integrals ranged from 0.768 to 0.814 and from 0.721 to 0.784 to distinguish benign from malignant tissue, respectively. Neither method displayed significant differences between measurements (automated vs. human expert readers, p > 0.05), in line with the results from the univariate Spearman's rank correlation coefficients, as well as the Passing and Bablok regression analysis. It was concluded that this pilot study demonstrates that 1H‐MRS data from breast MRI can be automatically exported and interpreted by SpecTec‐ULR 1.1 software. The diagnostic performance of this software was not inferior to human expert readers.

automated, biomedical image processing and analysis, breast/diagnostic imaging, breast/ pathology, data science, magnetic resonance imaging/methods, pattern recognition, proton magnetic resonance spectroscopy

| INTRODUCTION
3][4][5] However, the presence of contrast enhancement is not specific and ongoing research efforts aim to increase its specificity.Multiparametric MRI is defined as a combination of dynamic contrast-enhanced (DCE) T1-weighted, T2-weighted, and diffusion-weighted imaging (DWI) sequences, and it provides information on the vascular, microstructural, and biochemical properties of breast lesions.It can thus increase the specificity compared with contrast-enhanced images only. 6MR spectroscopy is a highly specific tool for diagnosis of breast cancer that can be implemented in breast MRI examinations, but is currently not clinically applied in routine protocols. 7,8This is despite the fact that MR spectroscopy provides unique additional information that can further tailor breast cancer diagnosis and prognosis. 9,10To date, MR spectroscopy is the only noninvasive molecular imaging method available in vivo.Total choline (tCho; resonance at 3.23 parts per million [ppm]) has been identified as a specific biomarker suitable for breast cancer diagnosis and characterization. 11Recently, the incremental value of single-voxel proton MR spectroscopy (1H-MRS) over "simple multiparametric" MRI protocols consisting of DCE, T2-weighted, and DWI, was documented. 8This value was mainly attributed to an increase in specificity while maintaining sensitivity.On the other hand, as opposed to, for example, DWI, 1H-MRS in the single-voxel technique is not able to directly visualize breast cancer lesions, requires more sophisticated planning and adjustments, and necessitates specific software tools for data analysis.Moreover, the basic chemistry/spectroscopy knowledge required to employ this method further limits its application in clinical practice. 7With the advent of high-field MRI and multichannel coil technology, the available MR signal for spectroscopy has increased and thus allows significantly shortened acquisition times of well within 3 min.
Yet, the main problem of complicated data analysis remains.We therefore tested software that automatically processes and analyzes body 1H-MRS spectra.3][14] The purpose of this study, therefore, was to test the performance of this software against human expert interpretation of breast 1H-MRS spectra using standard software in a prospective dataset.

| Patients
This study comprised 117 spectra of 74 consecutive patients.We retrospectively analyzed the data from patients who were enrolled in a prospective institutional review board-approved study and underwent breast MRI for evaluation of positive predictive value 1 (PPV1) recalls (breast imaging reporting and data system [BI-RADS] 0, 4/5 on conventional imaging) from June 2016 to March 2020.
The inclusion criteria were: (1) the patients had to be older than 18 years; (2) no prior neo-adjuvant chemotherapy or radiotherapy of the affected breast; and (3) written consent to participate in the study was provided.The exclusion criteria were: (1) prior neo-adjuvant chemotherapy or radiotherapy of the affected breast; (2) contraindications for MRI examinations; (3) images not suitable for subsequent analysis because of strong motion artifacts; (4) pregnancy or breast-feeding; and (5) no histopathology or long-term follow-up of more than 24 months was available.

| Cohort description
The analysis included 117 spectra of 74 patients that harbored 74 lesions.One spectrum was derived and analyzed per lesion.The remaining 43 spectra were derived from healthy contralateral sides of these patients.The patients were all females aged 32-84 (median 51) years.There were 28 benign and 46 malignant histologically verified lesions.The overall mean lesion diameter was 23.8 ± 21.4 (median 18, range 5-117) mm.
The malignant and benign mean lesion diameters were 22.7 ± 18.4 (median 18, range 5-87) and 27.7 ± 29.6 (median 16.5, range 5-117) mm, respectively.Malignant lesions comprised 20 invasive ductal carcinoma (IDC), 22 IDC with a ductal carcinoma in situ (DCIS) component, one DCIS, and three ILC.A benign diagnosis was established by 24 months of imaging follow-up.For more detailed information on patient and lesion characteristics, see Table 1.

| Reference standard
Histopathology served as the standard of reference and was obtained in suspicious lesions either by image-guided biopsy or surgery.The histopathological reference standard was established by an experienced board-certified breast histopathologist (with > 20 years of experience).In the absence of suspicious lesions, 24 months of imaging follow-up was mandatory.The reference standard was used to dichotomize the dataset into malignant and not malignant (i.e., benign) cases.

| Breast MRI and MR spectroscopy
Breast MR images, including 1H-MRS images, were acquired on a 3-T positron emission tomography (PET)-MRI scanner (Magnetom Biograph, Siemens, Erlangen, Germany) using a vendor-supplied dedicated 16-channel breast coil.Patients were examined in the prone position.The imaging protocol was in line with European Society of Breast Imaging recommendations and comprised an axial T2-weighted turbo spin echo sequence (TR 4820 ms, TE 192 ms, flip angle [FA] 128 , matrix 640 Â 480, field of view [FOV] 360 mm).For DCE-MRI, a T1-weighted Dixon Twist VIBE sequence was used (TR 4.7 ms, TE 1.3 ms, FA 10.5 , matrix 352 Â 352, FOV 440 Â 440 mm, 23 measurements at 14 s per acquisition).After two baseline scans, Gd-DOTA (Dotarem, Guerbet, France) was injected at a dosage of 0.2 mL/kg bodyweight, using an automated injector with 3 mL/s injection rate and including a final 20 mL flush of saline solution.The target lesion for single-voxel 1H-MRS using the Point-REsolved Spectroscopy Sequence (PRESS; TR 2000 ms, TE 135 ms, delta frequencyÀ2.5 ppm, vector size 1024, receiver bandwidth 1000 Hz, water suppression bandwidth 35 Hz, averages 64, acquisition time 2 min 16 s) was selected by the supervising radiologist based on all multiparametric images.In the case of several suspicious lesions, 1H-MRS was performed on the largest lesion.The spectroscopy voxel aimed to enclose the lesion while avoiding the inclusion of nonenhancing breast tissue and fat.Automatic first-and second-order shim gradient adjustment was applied to optimize static field (B0) homogeneity.In the case that a target full width at half maximum for water below 35 Hz could not be achieved automatically, manual shimming was performed.The sequence was repeated as an internal reference once with four averages (acquisition time 12 s) without water suppression, while otherwise retaining the exact adjustments.Overall, the 1H-MRS acquisition time including shimming was less than 5 min in all patients.

| Automated data analysis
Before the automated analysis, the spectroscopy data (water-suppressed und unsuppressed spectra of each patient) were manually exported as Siemens RDA files and thereupon imported into SpecTec-ULR 1.1 software (LifeTec Solutions GmbH), which is spectral data analysis software written in the Python programming language.The SpecTec-ULR 1.1 software is not fully open source but can be provided for scientific T A B L E 1 Clinical and histopathological properties of participants (n = 74).Abbreviations: G1, grade 1; G2, grade 2; G3, grade 3; HER2; human epidermal growth receptor 2; SD, standard deviation; TN, triple negative.
collaborations upon request.Note that an automated export pipeline from the MRI unit to the software does not require novel technological developments and mainly requires consent between the vendor and imaging facilities.
The software performs various postprocessing steps on the raw MR time-domain signal, such as zero filling to 2048 points to increase the spectral resolution and apodization by multiplication of the free induction decay (FID) signal with an exponential decay function with a time constant of 60 ms for noise reduction.
Further data processing operations to improve spectral quality are applied on the frequency-domain signal after Fourier transformation.
These operations include linear phase correction using the ACME algorithm 15 and baseline correction based on multipolynomial fitting and subtraction. 16fore single peak identification is initiated, the software analyzes the position of the water peak (and in relation to that the methylene peak where applicable) in the unsuppressed reference spectrum, to assure that the spectrometer operating frequency (scanner frequency) was tuned to the resonance frequency of water protons and hence the water peak is observed close to the center frequency (0 Hz).This represents the most frequent setting in NMR spectroscopy studies.In that case the center frequency (in Hz) is converted to the corresponding ppm value of 4.7 and an approximate adjustment of peak localization is ensured for further peak identification and fine-tuning of peak alignment.In the case of different scanner frequency settings, the frequency difference between the water peak and the center frequency is calculated and a first coarse frequency shift is carried out to move the water peak to the center of the spectrum and convert it to 4.7 ppm.
Peak identification is then performed by scanning the frequency ranges of olefinic acids (< 5.7 and > 5.1), methylene group (< 1.6 and > 1.0), methylene α-carbonyl group (< 2.5 and > 2.0), diallyl (< 2.9 and > 2.6), and water (< 5.0 and > 4.5) for maximum amplitude values in the watersuppressed spectrum.A peak alignment algorithm based on least-squares approximation is then used to shift the spectrum and align the peaks as close as possible to the known ppm values.No other alignment algorithms that adopt stretching or compressing methods, which could affect integration for quantification, are applied.
Once the peaks defined above are identified and the spectrum aligned so that the peaks are assigned as close as possible to their known frequency, other peaks such as tCho are identified by scanning the appropriate range in the spectrum (< 3.4 and > 3.0) for the maximum amplitude value of the flux density.Then the signal-to-noise ratio (SNR) is calculated as the ratio of the height (amplitude) of the peak and a combination of maximum amplitude and standard deviation values of the noise signal measured at defined areas at both ends of the spectrum (> 8 and < À1 ppm), as well as noise in the proximity of the choline peak.This is to account for possible baseline correction errors, especially in the case of baseline distortions at the extremities of the spectrum, which could significantly impact SNR calculations when they concern small peaks such as tCho.If the SNR of a single peak reaches a certain threshold (SNR > 2), spectral line fitting is performed, and the flux integral is calculated.
All 117 spectra were analyzed in this manner and the results (integral and SNR of the tCho peaks) were exported into a .csvfile for comparison with the results of the manual analysis and interpretation of the radiologists.The automated analysis was performed twice to ensure reproducibility.

| Manual data analysis
Two radiologists (PC, a consultant breast radiologist with > 8 years of experience in breast imaging and proton MR spectroscopy; and DK, a resident with 2 years of experience), who were blinded to the patients' history and histology results, independently performed a manual analysis and interpretation of all 117 spectra using the free-ware version of the commercial software iNMR v. 5.4.4 (www.inmr.net).The same postprocessing steps described in the previous section were manually applied.The precise steps were: zero filling of the FID signal to 2048 points, manual phase correction, and Gaussian apodization of 5 Hz of the time-domain signal.The water peak was assigned to 4.74 ppm as a reference, confirmed by using the unsuppressed spectrum.Positive Cho resonances were identified at 3.23 ppm.Finally, spectral quantification was performed by calculation of the area under the choline peak (integral), which required the user to manually define the constraints of the area for peak integration by the software.The SNR of tCho was calculated by taking the peak amplitude of tCho and dividing it by the amplitude of the noise, as measured on the left part of the spectra in the chemical shift range of 6.5-7.5 ppm.

| Statistics
Statistical analysis included the calculation of nonparametric Spearman's correlation coefficients for all measured parameters (SNR and integral of tCho, measured twice each by the SpecTec-ULR 1.1 software and two readers).Results were visualized using the corrplot routine in R. Subsequently, Passing and Bablok regression was used to assess systematic, proportional, and random differences between the averaged results of the SpecTec-ULR 1.1 software against both readers for SNR and integral tCho values.In Passing and Bablok regression, there is no need to distinguish between a reference and comparison method.Both methods can be considered interchangeable.Finally, the diagnostic performance to distinguish breast cancer from benign tissue was evaluated by receiver operating characteristics (ROC) analysis using the area under the ROC curve (AUC) as a general measure of diagnostic test performance.The results for each single reading were compared using the DeLong method.p values of 0.05 or less were considered statistically significant.No alpha error accumulation error correction was used for this exploratory analysis.All Passing and Bablok regression and ROC analysis was performed using Medcalc version 20 (Medcalc, Mariakerke, Belgium).

| Univariate Spearman's rank correlation coefficient analysis
The color-coded correlation matrix in Figure 1 shows that all tCho SNR and integrals measured by the novel software, as well as those interpreted by human expert readers, displayed a positive correlation coefficient (blue hues).All correlation coefficients (rho) were statistically significant ( p < 0.01).Software-extracted, as well as reader-extracted SNR and integral repetition, show very high correlation coefficients (0.97-1).If rho more than 0.7 is interpreted as strong correlation and rho more than 0.9 as very strong correlation, all univariately determined values in this analysis show at least a strong correlation.

| Bablok and Passing regression
The comparison of tCho SNR measurements demonstrated no systematic or proportional differences between software-extracted values and reader 1, while a minor systematic and proportional difference was present between software-extracted values and those obtained by reader 2. Between reader 1 and reader 2, a minor proportional difference was noted for tCho SNR values.In general, there was a tendency towards higher tCho SNR values measured by the software compared with both readers.Regarding tCho integrals, both readers showed a significant proportional difference to the software-extracted values that were found to be lower by the software, specifically in the case of high tCho integral values (Figure 2).
While there were no evident systematic differences between measurements, SNR values obtained by the software tended to be higher compared with those by human readers, indicating a certain proportional difference.This was true for the comparison with reader 2 (B).SNR measurements of reader 1 compared with reader 2 only showed minor differences (C).Moreover, there was quite a significant proportional difference between software-extracted tCho integral values that were lower compared with both readers (D and E), while both readers did not show systematic or proportional tCho integral differences (F).

| ROC analysis
The areas under the ROC curves for tCho SNR ranged from 0.768 to 0.814 (Table 2 and Figure 3) without significant differences between measurements ( p > 0.05, respectively).The AUC of tCho integrals ranged from 0.721 to 0.784 and also did not vary significantly between measurements ( p > 0.05, respectively).Figures 4 and 5 illustrate software output examples of a benign and a malignant lesion, respectively.

| DISCUSSION
The results from this pilot study show that the developed software is capable of automatically extracting tCho SNR and integral data similarly to human expert readers.Method reproducibility was excellent, yielding a 100% match of the derived data.When comparing the software performance with the human expert readers, we noted proportional differences between the human expert readers and the automated software,   This has significant implications: currently, 1H-MRS of the breast is only a niche application in multiparametric MRI because of the complexity of data acquisition and analysis.Although advances in MR technology, multichannel coils and high-field units, in the meantime allow for substantially reduced acquisition times without a significant loss of SNR or spectral quality, the necessity for human expert interaction to further process these data still limits its clinical applicability.More precisely, exporting raw spectral data to dedicated software, as well as expert knowledge necessary for data evaluation and interpretation, are currently time consuming and dependent on spectroscopy training, which only very few select clinical radiologists possess.1H-MRS was developed for brain applications where a relatively high signal for standard diagnostic peaks is observed upon 1H-MRS.The tCho peak, on the other hand, often shows a low SNR, which was on average 2.042 (95% CI 1.42-2.65) in this study, explaining difficulties in automated quantitation.On the other hand, 1H-MRS has already been proven to provide unique information with important clinical and prognostic implications. 8The automated workflow used in this software allows for direct scanner integration and is also able to provide batch analyses required for large-scale in vivo biomarker research.This software could resolve the issue of time-consuming data export and analysis, as well as reduce the need for dedicated spectroscopy training, thus enabling implementation of 1H-MRS in clinical routines.
With our automated software approach, employing 1H-MRS in clinical practice becomes feasible.It still requires further research, but the basic process appears sound, given the data from this pilot study, opening the door to a new era in breast MR imaging.
8][19] Yet, Section 2 unveils a lack of standardized tools and that the assessment of tCho is a time-consuming process that requires regular intervention by a human reader with high expertise, illustrating the clear need for automated quantitation tools.Simultaneously, studies on the reproducibility of such automated quantitation tools are a "conditio sine qua non" for any future application.
This study has several limitations.First and foremost, this is only a pilot study, employing an experimental setting to test the reliability of the software against human readers.Further studies exploring the diagnostic performance of this automated algorithm, ideally against human expert readers, should be carried out in a prospective multicentric trial better reflecting different clinical conditions, to ensure robustness and clinical validity.Second, we only compared our automated approach with human readers relying on one given (user-dependent) software.Future studies could investigate human reader performance relying on an array of different (user-dependent) software.In line with other reports, the SNR of water in our study was generally very high compared with the SNR of choline, which is often low.The former most likely reflects the increased cell density, vasculature, and inflammatory reaction, along with decreased gross fat content, observed in malignant breast lesions.By comparison, the SNR of choline can sometimes present with low levels, especially in diffuse growing pathological entities such as ILC and DCIS, probably attributable to the low tumor cell density and proliferative activity in these entities. 18The low signal of tCho poses a quantification problem for commercial in vivo spectroscopy analysis tools, and measuring the tCho SNR in these lesions can potentially compensate for this.Our clinical dataset was acquired using only 64 averages, which was necessary to integrate the spectral data acquisition in clinical practice.This probably has a detrimental effect on sensitivity because of a reduced SNR of an a priori low SNR peak such as tCho.Another factor limiting tCho SNR is the spectral data acquisition after IV contrast.Particularly ionic contrast media (as used in this study) have a broadening effect on peak width and thus SNR is reduced. 20To date, the effect on other metabolites including lipids has been insufficiently investigated.However, both aspects do affect both manual and automated readout and do not limit the applicability of our results as no systematic bias is conceivable.Low SNR may let both automated and manual analysis miss low SNR peaks such as tCho and limit the sensitivity of the method.A too aggressive baseline correction may add to this effect.This is reflected in our results, which show a minor reduction in sensitivity of the automated tCho integration compared with the human reader approach that used an individually set baseline correction, which was not statistically significant.We did not investigate whether the software works with spectra derived from different MRI units, vendors, and sequence protocols, as the software can consistently extract spectroscopic data akin to a human reader, irrespective of the data source (i.e., MRI machine, acquisition parameters).There is no inherent reason, either from an MR physical or data-processing perspective, that would hinder our software from analyzing a single-voxel spectrum from any system.The potential challenge of data compatibility is simply a technicality that can be solved prior to software application.

| CONCLUSION AND FUTURE OUTLOOK
We tested automated software, capable of exporting and interpreting 1H-MRS data from breast MRI.Compared with human expert readers, the diagnostic performance of this software was not inferior in this pilot study.Our findings have widespread applications: because of its automated approach, spectroscopy analysis could be integrated in the data pipeline between the MRI scanner and the picture archiving and communication system.Radiologists could directly use the diagnostic and prognostic information provided by 1H-MRS clinically.In addition, batch analyses of large amounts of data could be analyzed and used for testing hypothesis and building 1H-MRS-based decision support systems ("in vivo metabolomics").

1
Correlation plot of tCho SNR and integrals measured, using the developed software and the human readers.The correlation coefficients (all significant at p < 0.01) are given in the lower left triangle and are represented graphically within the upper right triangle.A colorlookup table is provided on the right; positive correlations are marked in blue, while negative correlations are marked in red.The key is (all referring to tCho): (a) software-extracted SNR repetition 1; (b) software-extracted SNR repetition 2; (c) SNR reader 1 repetition 1; (d) SNR reader 1 repetition 2; (e) SNR reader 2 repetition 1; (f) SNR reader 2 repetition 2; (g) software-extracted integral repetition 1; (h) software-extracted integral repetition 2; (i) reader 1 measured integral repetition 1; (j) reader 1 measured integral repetition 2; (k) reader 2 measured integral repetition 1; and (l) reader 2 measured integral repetition 2. SNR, signal-to-noise ratio; tCho, total choline.

F
I G U R E 3 ROC curves using SNR and integral values of tCho as measured by the developed automated software, reader 1 (R1), and reader 2 (R2) to distinguish benign from malignant breast tissue.Note that only one curve for each automated software-measured tCho SNR and integral is given, because the repetition yielded exactly the same results.ROC, receiver operating characteristics; SNR, signal-to-noise ratio; tCho, total choline.F I G U R E 4 Exemplary software output of a benign lesion.(A)The water peak is identified on the nonwater-suppressed reference spectrum (blue vertical line).(B) Full spectrum with water suppression and the successful identification of the tCho area at 3.23 ppm and the control peaks at 5.34, 2.81, 2.3, and 1.33 ppm.(C) Magnification of the analyzed tCho area in the full spectrum.The SNR of tCho is 1.Histopathology revealed a benign finding.ppm, parts per million; SNR, signal-to-noise ratio; tCho, total choline.F I G U R E 5 Exemplary software output of a malignant lesion.(A)The water peak is identified on the nonwater-suppressed reference spectrum (blue vertical line).(B) Full spectrum with water suppression and the successful identification of the tCho area at 3.23 ppm and the control peaks at 5.34, 2.3, and 1.33 ppm.(C) Magnification of the analyzed tCho area in the full spectrum.The SNR of tCho is 19.Histopathology revealed an invasive G3 breast cancer.G3, grade 3; ppm, parts per million; SNR, signal-to-noise ratio; tCho, total choline.specifically regarding tCho integral values.Yet and importantly, the diagnostic performance, as quantified by the area under the ROC curve, did not differ.