2.1.

System Overview

The complete TOBI2 system, together with its schematic drawing, is shown in Figs. 1 and 2, respectively. The optical probes are directly attached to the Hologic Selenia dimensions DBT machine and connected via glass fibers to the instrument tower. The tower houses both the RF and CW subsystems. Due to the necessity to locate TOBI2 in an active clinical space, the tower is fully enclosed. On the outside of the tower, there is a shelf with the laptop controlling the optical system, as well as holders for the optical probe allowing storage when the DBT machine is used alone. In the following paragraphs, we describe each component in more detail.

Fig. 1

(a) Complete TOBI2 system. Optical probe is attached to the DBT machine. Optical fiber bundles connect the optodes in the probe to the instruments inside the enclosed tower. (b) Inside view of the instrument tower. From top to bottom: the RF detectors, RF sources, ($2\times$) CW source expansion boxes, and the CW6 main instrument. (c) Close-up view of the source plate. (d) Close-up view of the detector plate.

Fig. 2

Schematic overview of the TOBI2 system. CW components are shaded in yellow, RF components are shaded in light green, DBT system components are shaded in gray, and fiber optics are represented in blue.

2.2.

CW Component

The CW imaging unit, shown in Fig. 1(b), was manufactured by TechEn Inc. (Milford, Massachusetts). It consists of the main CW6 system, containing 32 detectors and 32 lasers diodes split evenly between the wavelengths of 690 and 830 nm. The unit also contains two supplemental boxes, each containing 32 lasers split again evenly between 690 and 830 nm, for a total of 96 CW lasers. Each laser is modulated with a square wave at one of 32 discrete frequencies between 6.4 and 12.6 kHz. The frequencies are chosen to span less than one octave, so that harmonics do not fall on other modulation frequencies. The system powers the lasers in the main CW6 unit and the supplemental boxes in a sequential, electronically switched order, so that only 32 lasers are on simultaneously. The dwell time in each state is specified in multiples of 40 ms, and for the data shown in this paper, we used a dwell time of 1 s, thus resulting in a frame rate of $1/3\mathrm{Hz}$.

On the detection side, each channel consists of a Hamamatsu avalanche photo diode (APD) module (C5460-01) followed by programmable signal amplification, conditioning, and digitization. On board field programmable gate arrays (FPGAs) and digital signal processors demodulate the signals by calculating the fast Fourier transform (FFT) of each detection channel and sending the raw intensity data of the appropriate FFT bin to the controlling computer via a universal serial bus (USB) connection at a rate of $25\text{samples}/\mathrm{s}$.

2.3.

Radio Frequency Component

The RF imaging unit, shown in Fig. 1(b), was built at the Martinos Center in collaboration with TechEn Inc. and was described before in Ref. 61. It contains one 685-nm laser diode modulated at 67.5 MHz, and one 830-nm laser diode modulated at 75 MHz. Both diodes are set to an average power of $\sim 25\mathrm{mW}$ and a modulation depth of 90%. The light of these lasers is collimated and then combined with a dichroic mirror. The resulting dual wavelength beam is launched via a two-dimensional galvo into one of 24 fibers. In the data shown here, the dwell time at each location is set to 0.13 s for the phantom measurements and 0.4 s for the patient measurement, resulting in cycle times over all 24 locations of 3.3 and 10 s, respectively.

On the detector side, each of the 20 channels consists of a Hamamatsu APD module (C5331-04) followed by amplification and filtering. The signal of each channel is directly digitized at 180 million samples/s without down conversion by a 16-bit analog to digital converter (ADC). Each ADC has an FPGA attached that computes overlapping 4 million point discrete Fourier transforms to demodulate the signals, resulting in a final data rate of 90 Hz per wavelength. The data from all detector channels are collected and sent to a computer via USB by a control card.

2.4.

Optical Probe

The optical probe consists of a source plate, shown in Fig. 1(c), which is permanently attached to the compression paddle of the DBT machine, and a detector plate, shown in Fig. 1(d), which fits on the x-ray detector cover of the DBT machine. The source plate contains 120 $500\text{-}\mu \mathrm{m}$ poly (methyl methacrylate) (PMMA) fibers, placed individually in milled channels in a 1/8-in. thick clear polycarbonate plate. At the edge of the source plate, the PMMA fibers are coupled into $500\text{-}\mu \mathrm{m}$ glass fibers, which transport the source light from the CW and RF imaging units. At the light emitting ends, the PMMA fibers are polished at 45-deg angles to send light into the breast tissue. The fiber channels end within 5-mm black plastic disks to prevent light leakage.

The detector plate consists of a quarter inch thick black acrylonitrile butadiene styrene (ABS) plate, into which 54 channels containing 2-mm PMMA fibers are milled. On the light collection side, the fibers are terminated with plastic prisms, and at the edge of the detector plate they are coupled into 2.5-mm diameter glass fiber bundles that route the light back into the instrument tower and to the detectors.

The optode locations on the source and detector plates were chosen to give full breast coverage in 80% of patients, as determined by breast outlines obtained from DBT images from our previous study.⁴⁹ Assignment of the source and detector locations to the CW and RF subsystems as well as to a specific wavelength has then been optimized using a genetic algorithm. Both the source and detector plates do not contain any glass or metal parts within the field of view of the DBT system to achieve x-ray translucency and minimize x-ray contrast.

2.5.

Phantom

To measure the temporal and spatial resolution of TOBI2, we created two phantoms. The first phantom features one centrally located spherical cavity with a diameter of 19 mm. The second phantom features three spherical cavities of 13-, 19-, and 25-mm diameter separated by 45 mm. The cavities can be filled with liquids of varying optical properties through the channels embedded in the phantom to either match with or to create contrast to background optical properties for imaging.

Each phantom was constructed from 2.8 L of silicone (Smooth-On Ecoflex 00-50) mixed with 2660 mg of white pigment and 78 mg of black pigment (Smooth-On Silc Pig White & Black). To avoid light piping, the inclusions were made without glass or any other material that could distort our results. Specifically, as seen in Fig. 3(a), the inclusions were made of water-soluble wax spheres (Freeman Sol-U-Carv) and were mounted in the mold with 2.4-mm diameter steel tubes. After pouring and curing of the degassed silicone, the tubes were removed, and the wax spheres were dissolved by flowing warm water through the channels. The finished phantoms, as shown in Fig. 3(b), have a thickness of 52 mm. The size and separation of inclusions were further confirmed by DBT x-ray control images, shown in Fig. 3(c) for the triple-inclusion phantom.

Fig. 3

(a) Mold for the triple-inclusion phantom, with wax balls forming the inclusions mounted on steel tubes. (b) Finished triple-inclusion phantom, with glued-on tubes to fill inclusions with liquid. (c) DBT slice of finished triple-inclusion phantom showing the three spherical cavities with connecting channels.

2.6.

Patient Imaging

A 47-year-old non-Hispanic white female with a breast cancer diagnosis was imaged on our system. An ultrasound guided left breast core biopsy indicated the presence of grade 3 invasive ductal carcinoma at the 5 o’clock position, located 3 cm from the nipple, measuring $1.8\times 1.2\times 1.1{\mathrm{cm}}^3$. An axillary lymph node core biopsy found lymph nodes with metastatic ductal carcinoma. Patient consent was obtained in accordance with the policies and guidelines of the Massachusetts General Hospital/Partners Healthcare Institutional Review Board. The subject’s breast was imaged under partial compression (half mammographic force, 21.8 N for this patient) first, followed by imaging under full mammographic compression (44.5 N for this patient). The imaging session lasted $\sim 3\min$.

2.7.

Image Reconstruction Methods

Optical image reconstructions are performed on a pair of tetrahedral meshes, a finer one for solving the optical forward problem and a coarser one for solving the inversion, generated using the MATLAB-based meshing toolbox “iso2mesh.”⁶² A slab geometry is used to generate phantom meshes, whereas 3-D DBT images are used to extract the breast shape for patient scans. Raw optical measurements are first calibrated against a homogenous phantom with known optical properties and then fitted for bulk properties. Nonlinear, spectrally constrained inversion of the finite-element representation of the diffusion approximation using the Tikhonov-regulated Gauss–Newton approach is performed for nine iterations using our in-house software, i.e., Redbird,⁴⁸ to reconstruct optical images shown in this paper. When solving the inverse problem, compositional structural priors are used as soft constraints in our structural-prior guided reconstruction algorithm described previously.⁶³ In phantoms, a sphere located at the inclusion center, and with its diameter matched, is used to derive the structural prior for each phantom inclusion, and the prior for background is set to enforce unity of both tissue compositions on each mesh node. For patients, a dual-Gaussian segmentation algorithm⁶³ is used to automatically derive adipose and fibroglandular compositional priors from DBT images. Similar to the phantom inclusion prior, a Gaussian sphere profile is used to generate an additional lesion prior with known centroid and size information provided by an experienced radiologist (Saksena).

3.1.

System Characterization

The RF component has already been characterized in Ref. 61. Here we will briefly summarize these findings. We measured a noise equivalent power of less than $1.4\mathrm{pW}/\surd \mathrm{Hz}$, which approaches the manufacturer specified noise floor of the APD module of $0.8\mathrm{pW}/\surd \mathrm{Hz}$. Channel separation on the detection side is greater than 100 dB (20 log 10); however, on the source side due to the construction of the galvo multiplexer, it is only 80 dB (20 log 10). Saturation is reached with an input signal of $1.5\mu \mathrm{W}$, thus combined with the noise equivalent power mentioned above, we claim a dynamic range of 115 dB (20 log 10). We were not able to observe any interwavelength or amplitude to phase crosstalk. The phase noise of the output signal is smaller than $6\mathrm{mrad}/\surd \mathrm{Hz}$ at 100-pW input power. Over 10 h, the measured amplitude changes by less than 1.5%, the phase less than 3 mrad at an optical power of 5 nW. We also characterized the CW component (CW6) using the same multiattenuation filter wheel setup as used in Ref. 61 for the RF component. We measured an instantaneous dynamic range of more than 103 dB (20 log 10) and a noise floor of $0.06\mathrm{pW}/\surd \mathrm{Hz}$.

To quantify the additional losses incurred due to choosing an x-ray translucent optical probe design, we measured the light transmission of both the source as well as the detector fibers and compared overall throughput versus a simple design where glass fibers directly touch the tissue. On the source side, power is reduced by 3.3 dB, and on the detection side by 6.4 dB at 830 nm. At 690 nm, the losses are slightly less due to the better transmissivity of the PMMA fibers at this wavelength. The detailed results are shown in Fig. 4. DBT slices of a patient’s breast taken with the optical probes in place can be seen in Figs. 5(a)–5(c). In Fig. 5(a), on the upper surface of the breast, the small source fibers are only barely visible, whereas in Fig. 5(b), on the lower surface of the breast, the larger detector fibers and prisms are clearly visible, but their artifacts do not exceed the dynamic range of the x-ray detector. Figure 5(c) shows a center slice, which can be compared to a center slice of the same patient taken without the optical probe attached [Fig. 5(d)]. It is evident that in the center of the breast the DBT reconstruction algorithm already largely removes the artifacts.

Fig. 4

Schematic showing the additional optical losses incurred due to the hybrid PMMA/glass fiber design versus a minimal fiber-to-tissue approach.

Fig. 5

Impact of optical plates on the DBT images. (a) Top slice of breast DBT volume taken with the optical probe attached. The source fibers and mounting holes are clearly visible. (b) Bottom slice of breast DBT volume taken with the optical probes attached. The detector fibers and prisms are clearly visible. The high absorption patches are the result of small pieces of electrical tape used in the construction of the probe. They were removed subsequently. (c) Middle slice of breast DBT image taken with the optical probe attached. Faint shadows of the detector fibers can be seen. (d) For comparison, middle slice of the separately acquired clinical breast DBT image taken on the same patient (no optical components present).

3.2.

Static and Dynamic Phantoms

To test our complete system with all its components, we performed a series of phantom experiments. As a first test, to determine the maximal useable source–detector separation, we measured the single-inclusion phantom described in Sec. 2.5 with the liquid inside the inclusion matched to the bulk optical properties, which in turn are comparable to those of a typical human breast (${\mu }_a=0.075{\mathrm{cm}}^{-1}$ at 690 nm and $0.052{\mathrm{cm}}^{-1}$ at 830 nm, ${\mu }_s^{\prime }=8.4{\mathrm{cm}}^{-1}$ at 690 nm and $7.1{\mathrm{cm}}^{-1}$ at 830 nm). The resulting calibrated signal magnitudes versus source–detector separations can be seen in Fig. 6. From this figure, we observe that separations of less than $\sim 9\mathrm{cm}$ result in detectable signals above the noise floor, whereas instrument noise dominates in measurements with source–detector separations of more than 9 cm.

Fig. 6

Plot of the signal amplitude of all possible source–detector combinations versus the corresponding source–detector distances measured in the single inclusion silicone phantom. The optical properties of the inclusion are matched to the bulk, which in turn has properties comparable to a human breast.

To demonstrate the image reconstruction algorithm and evaluate the spatial resolution, we also imaged the triple-inclusion phantom described in Sec. 2.5. All three inclusions were filled with a water, milk, and India ink mixture, with the scattering coefficient matching the bulk of the phantom and the absorption coefficient being 1.81 times the value of the bulk at 690 nm. Figure 7(a) shows an absorption image representing the middle slice of the reconstructed 3-D absorption map. The three inclusions can easily be seen, and the centroids are at the expected locations. The reconstructed absorption values at 690 nm are 0.138, 0.142, and $0.146{\mathrm{cm}}^{-1}$, which represent contrasts of 1.59, 1.63, and 1.68, respectively, to the reconstructed bulk absorption coefficient of $0.087{\mathrm{cm}}^{-1}$ at 690 nm.

Fig. 7

(a) Middle slice showing the reconstructed absorption map of the triple-inclusion phantom. (b) Middle slice showing the reconstructed absorption map of the single-inclusion phantom during the last third of the measurement. The blue and red markers indicate the location of the corresponding measurements plotted in (c). (c) Time course of the reconstructed absorption coefficients of the single-inclusion phantom at the location of the inclusion (blue), and at a control location (red).

To also test the temporal dynamic performance of the system, the single-inclusion phantom was imaged over 165 s, with one image being reconstructed for every 3 s of data. During the first third of the measurement period, the water, milk, and India ink mixture in the inclusion was matched to the bulk optical properties both in absorption and scattering. During the middle third of the measurement, the absorption was increased to $1.54\times$ the baseline value by injecting a different liquid mixture with higher ink concentration into the inclusion. For the last third of the measurement period, the absorption was further increased to $2.82\times$ of the baseline value. Figure 7(b) shows an absorption image representing the middle slice of the reconstructed 3-D absorption map during the last third of the measurement. Figure 7(c) shows the time course of the reconstructed absorption values at the center of the inclusion, as well as at the indicated location away from the inclusion for control purposes. The reconstructed absorption value at the control location has an average of $0.08{\mathrm{cm}}^{-1}$ and stays constant within $\pm 6\%$. The average values of the reconstructed absorption coefficients at the inclusion centroid during each third of the measurement are 0.072, 0.108, and $0.166{\mathrm{cm}}^{-1}$, respectively. Compared to the targeted $1.54\times$ and $2.82\times$ contrast, the reconstructed contrast during the middle and last thirds of the dynamic measurement are $1.5\times$ and $2.31\times$, respectively. The two visible spikes are presumably due to an expansion of the cavity due to increased liquid pressure when changing the mixture.

3.3.

Patient Images

Figures 8(a) and 8(b) show the slice of the total hemoglobin (HbT) concentration map passing through the center of the lesion, overlaid on the corresponding slice of the DBT from the patient scan. Optical images were reconstructed using the adipose and fibro-glandular tissue fractions as priors derived from the x-ray information.⁶³ In addition, we used a Gaussian sphere tumor prior with a diameter of 15 mm at the centroid of the lesion,⁶³ as determined by our collaborating radiologist (Saksena). Figure 8(a) shows the absolute HbT concentration and Fig. 8(b) shows the change in HbT due to increasing the compression level from partial to full mammographic force. To facilitate the visibility of hemodynamic changes in the tumor, the color scale in Fig. 8(a) is chosen such that values below $25\mu \mathrm{M}$, which are representative of normal tissues, are transparent. Similarly, the color scale in Fig. 8(b) is chosen to show positive HbT changes as transparent. In both images, the tumor area displays localized contrast, increased HbT concentration in the absolute image, and a compression-induced further reduction in HbT in the relative changes image. Figure 8(c) shows the time course of the mean HbT values in both the tumor region, and in the rest of the breast, respectively. During the first (half-force) compression period, HbT displays a slowly increasing trend in both the tumor and the normal tissues, but the tumor area exhibits what appear to be blood volume oscillations that are not as evident in the normal tissue. The second compression period was rather short, and the distinguishing feature is the substantial decrease in tumor HbT versus half-compression while only a small further decrease occurs in the normal tissues.

Fig. 8

(a) Absolute HbT concentration map of the slice corresponding to the tumor centroid, overlaid over the corresponding x-ray DBT slice. (b) Change in HbT concentration as the compression is increased from half to full mammographic force. The red line in (a) and (b) denotes the tumor outline as marked by our collaborating radiologist. (c) The time course of HbT concentration in the tumor region and normal breast tissue, respectively, during the entire measurement session (the vertical bar and break in the timeline indicate where compression was increased from half to full mammographic compression).