2.1.

Cherenkov Imaging

An intensified, gated, complementary metal oxide semiconductor (iCMOS) camera (C-Dose, DoseOptics LLC, Lebanon, New Hampshire) equipped with an AF Nikkon 50 mm f/2.8 lens (Nikon Inc., New York) was mounted on the ceiling at IEC 61217 spatial coordinates of ($-1288$, 1066, and $-687$) mm from the linac isocenter. The camera captured Cherenkov emission from each homogeneous phantom throughout 200 MU of a 6 MV external beam delivery, or each patient throughout their respective treatment plan. Gated operation of the image intensifier, which was enabled only throughout the duration of $4\text{-}\mu \mathrm{s}$ x-ray pulses of the accelerator, allowed rejection of the room light signal in between each pulse. Each frame had a fixed exposure time of 51 ms, containing up to 18 linac pulses (in-sync frame), and was immediately followed by acquisition of an out-of-sync frame with 8 ms exposure time. During acquisition of the background, or out-of-sync frame, the intensifier was enabled over 720 ms with a 100-ms delay after the linac pulse, ensuring that only room light is recorded. Interleaved in-sync and out-of-sync frame acquisition enabled real-time x-ray noise filtering and room-light background subtraction, producing a clean Cherenkov emission video overlaid on live background video [Fig. 1(b)]. For each delivery fraction, the background-subtracted images were temporally summed across the duration of beam delivery, producing a cumulative CI that is independent of linac dose-rate fluctuations.

2.2.

Spatial Frequency Domain Imaging

An SFDI system (Reflect RS, Modulated Imaging Inc., Irvine, California) was used to recover the optical properties of phantoms and patient tissue. SFDI fringe patterns with varying spatial frequencies (0.0, 0.05, 0.1, 0.15, and $0.2{\mathrm{mm}}^{-1}$) over three phases (0 deg, 120 deg, and 240 deg) were projected onto the target and imaged at 659, 691, 731, and 851 nm wavelengths. This system was customized to operate at a working distance of 42 cm and has field of view of $29\mathrm{cm}\times 22\mathrm{cm}$. A sample image set of each of five spatial frequencies (one phase, one wavelength) is depicted in Fig. 2(a). The image was then acquired, and reflectance data were calibrated using a phantom of known optical and reflectance properties, yielding a calibrated reflectance $R_d$ for each wavelength at each pixel. The calibrated reflectance also includes a correction of surface topography based on previously published algorithms.¹¹ The calibrated reflectance value is evaluated by the internal reflect RS homogenous Monte Carlo-based multiple-frequency lookup table (LUT)-based fitting algorithm to return corresponding ${\mu }_a$ and ${\mu }_s^{\prime }$ [Fig. 2(b)].¹² The best fit will return a combination of ${\mu }_a$ and ${\mu }_s^{\prime }$. The ${\mu }_a$ and ${\mu }_s^{\prime }$ maps, such as those in Fig. 2(c), were generated with $520\times 696\text{pixel}$ resolution.

Fig. 2

(a) One planar image and each of four spatial frequency images are taken at phases 0 deg, 120 deg, and 240 deg, per wavelength. This process is repeated for four wavelengths, where modulated images from only one phase and one wavelength is shown. (b) The LUT takes diffuse reflectance measurements and calibrates each reading based on an included phantom with known optical and reflectance properties. Reflectance is calibrated from both the planar frequency $f_x=0{\mathrm{mm}}^{-1}$, and one other spatial frequency, this pixel is matched with a specific absorption and reduced scatter pair using an LUT much like that pictured. (c) Optical properties are fit for each pixel ($520\times 696$), forming the images pictured.¹⁰

Beginning with homogeneous phantoms, the SFD image was taken normal to the phantom surface to ensure focus across the entire surface of the phantom and to minimize artifacts due to phase unwrapping in height correction process. Each image took $\sim 30\mathrm{s}$ to acquire, followed by several minutes to process for optical property maps. The SFDI system was always removed from close proximity of the linear accelerator prior beam delivery. All Cherenkov and SFD image postprocessing and analysis were done in a MATLAB environment (Mathworks, Natick, Massachusetts).

2.3.

Tissue Phantoms

Tissue-equivalent phantoms were used prior to patient imaging to confirm the range of intensities expected in both CIs and SFDI images and to generate the tissue optical property calibration curve.¹³ The phantoms consisted of water, Intralipid (soybean and egg yolk fat emulsion, Baxter Healthcare Corporation; Deerfield, Illinois), and blood (bovine whole blood in Na Heparin, Lampire Biological Laboratories; Pipersville, Pennsylvania), where blood concentrations were varied from 5% to 23%, and Intralipid remained fixed at 1% (diluted from 20%) and added at each iteration to compensate for increasing volume due to added blood. This equated to an effective attenuation coefficient range between 0.2 and $0.45{\mathrm{mm}}^{-1}$. The blood/Intralipid suspension was contained in a 5-cm diameter, 1-cm deep plate. A stir plate and pellet were used to ensure medium homogeneity for nonvessel tissue phantoms. The phantom container was painted matte black to reduce glare artifacts that arise during both Cherenkov and SFDI imaging.

In a second study, we validated and assessed the limits of our technique applied to an optically heterogeneous target using a range of synthetic vessel phantoms. The blood vessels were simulated using clear polytetrafluorethylene (PTFE) tubing with 100% blood, diameters of 0.8, 1.9, and 2.3 mm and respective wall thicknesses of 0.46, 0.3, and 0.3 mm. The tubing was fixed in a black-coated well plate with a diameter of 14 cm and secured with a sealing glue. The surrounding media was constituted by a 0.5% blood and 1% Intralipid suspension to mimic properties of adipose tissue.¹⁰ These phantoms contained an adjacent length of tubing parallel to the tubing containing blood, $\sim 2\mathrm{cm}$ to the left, filled with bulk media. These measures were carried out to ensure that the amount of Cherenkov light coming from the tubing did not compromise the integrity of the signal from the sample. This was the case for the first selection of tubing, which led to the selection of thinner (PTFE) tubing that incited far fewer (if not negligible) Cherenkov photons upon irradiation. The well plate provided a sufficient volume of media beneath the tubing to incite underlying Cherenkov emission from 5 mm below the vessel, and the starting measurement was taken with enough media to cover the surface of the tubing. An SFD image was taken for this phantom, then $\sim 15.2\mathrm{mL}$ of media was added to superimpose 1 mm of media on top of the vessel, further submerging it. This process continued every 1 mm until the vessel was submerged beneath 5 mm of media. Manual mixing at each step ensured homogeneity of the phantom surrounding the vessel. After completing SFD imaging, the superimposed media was removed and this process was repeated for the same series of Cherenkov acquisitions, ensuring that the placement of the imaging system and the sample remained the same between additions of media, facilitating a robust image registration.

2.4.

Patient Radiation Therapy

All human imaging was done under an approved Institutional Review Board (IRB) protocol at Dartmouth College, and all procedures were carried out as described in this protocol. The recruitment to participate included informed consent for imaging the treatment area before and during radiotherapy. All subjects were treated by their clinically prescribed whole breast radiotherapy plan using beams from angles optimized to minimize damage to healthy tissue. These include right posterior (RP), left anterior (LA), right posterior oblique (RPO), and left anterior oblique (LAO) beam angles, prior to or following breast conserving surgery. Patient 1 was prescribed a hyperfractionated plan, 5 fraction/week treatment over 6 weeks, 4 beams centered around the breast, LAO (47 cGy), LA (50 cGy), RPO (42 cGy), RP (41 cGy), and two superclavicular fields, RPO (152 cGy), and LAO (28 cGy). Patient 2 was prescribed a standard hypofractionated plan, 5 fraction/week treatment over 4 weeks, which included a delivery of two beam angles, RPO (137 cGy) and LAO (129 cGy) at each fraction by a Varian CD2100 Linac (Varian Medical Systems Inc., Palo Alto, California), incorporating both 6 and 10 MV beams in most patients. All treatments were carried out at the Norris Cotton Cancer Center, a component of Dartmouth-Hitchcock Medical Center (Lebanon, New Hampshire).

2.5.

Theory

The empirical foundations upon which this study is based rely on, namely, the validity of diffusion theory in homogeneous media and Beer’s Law. We construct a tissue optical property correction technique in the general form of $I_0=I/\mathrm{CF}({\mu }_{\mathrm{eff}})$, where $I$ is the detected or observed Cherenkov light intensity, $I_0$ is the corrected light intensity, and $\mathrm{CF}({\mu }_{\mathrm{eff}})$ is a correction function, dependent on the effective attenuation coefficient. It is logical to imply that this correction function must also depend on tissue heterogeneity and on Cherenkov source distribution. For clarity and due to the limitations of our imaging techniques, we avoid resolving the Cherenkov source depth distribution by introducing an average sampling depth $d$ and assume that the tissue is optically homogeneous in depth.

To facilitate clinical tissue optical property correction, a calibration curve was constructed from the relationship between detected Cherenkov emission and measured effective attenuation. From the reduced scatter and absorption maps provided by the SFDI system, the effective attenuation coefficient maps were generated, given the approximation in diffusion theory

The exponential decay relationship in consideration is illustrated by the Beer–Lambert law:

For each pair of corrected and uncorrected images, the contrast was computed by taking the mean intensity value of an ROI inside the vessel over the mean intensity of an ROI of the surrounding bulk media. This metric is used to ensure consistency across calculations by normalizing to the conditions of the surrounding media environment. The same procedure is carried out for in vivo vessels in patient images with an additional step, which includes comparison to the treatment plan, such that the expected dose gradient of the plan is considered. Finally, all images are coregistered to ensure the same ROI is being evaluated.

3.1.

Homogeneous Phantom Calibration Curve

Highly absorbing features, such as blood and skin pigment, have a vast influence over observed Cherenkov intensity, illustrating reductions up to 60% in the calibration data set. In Fig. 3(a), a white light image of each homogeneous phantom is shown over the series of increasing blood concentrations (top row). Below, the respective effective attenuation map [Fig. 3(b)] and CI [Fig. 3(c)] is recovered for 851 nm central wavelength. As effective attenuation increases (indicated by color bar, left), the fluence of the Cherenkov emission decreases (color bar, right). The mean values of measured Cherenkov optical intensity and effective attenuation coefficients were used to construct the calibration curve in Fig. 4(a), by way of Eq. (4). This data set was fit using each of the four NIR wavelengths used by the SFD imaging system, where the calibration at 851 nm illustrated the best fit and was thus chosen for application. This confirmed an exponentially decreasing emission of Cherenkov light with an increasing blood concentration, and therefore, ${\mu }_a$ and ${\mu }_{\mathrm{eff}}$. To demonstrate this proof of this concept, the fit used to determine the blood–Intralipid phantom calibration is applied back to the same data set, to illustrate the extent of Cherenkov emission correction in an ideal case [Fig. 4(b)]. The corrected Cherenkov intensities approach values that would be measured in a phantom with virtually no absorption. In this particular case, we observed a mean corrected intensity of $\mu =3.8\times {10}^4$ counts, with a standard deviation of $\sigma =879$ counts. The coefficient of variation ($\mathrm{CV}=\sigma /\mu$) equaled $\sim 2.3\%$. (Note this $\mu$ as an average and is not to be mistaken for an attenuation coefficient.)

Fig. 3

(a) An observable increase in effective attenuation follows a series of 5% to 24% blood concentrations (first row) where blood is a known optical light absorber. (b) SFDI absorption and scatter maps were taken for each sample and used to create effective attenuation maps using the expression in Eq. (1). (c) The respective CIs illustrate the reduction in Cherenkov light with increasing ${\mu }_{\mathrm{eff}}$. (d) Once the correction has been applied, the average value of each well equalizes to $3.81\times 104$ counts with a standard deviation of 879 counts, $CV$ $(\sigma /\mu )=0.023$. The decreased Cherenkov intensity in the center of each well is an artifact due to the stir pellet.

Fig. 4

(a) The calibration data set between SDFI-measured effective attenuation coefficients (in vivo range, 851 nm) and the associated average detected Cherenkov outputs (natural log scale). The measured Cherenkov emission $\ln (C)$ and effective attenuation (${\mu }_{\mathrm{eff}}$) pair are used to extrapolate the intercept $C_0$ or $C_{\text{corrected}}$, where $C_{\text{corrected}}$ and effective sampling depth d have been fit linearly. (b) The uncorrected average ROI values of each phantom (blue) illustrate exponential decay with increasing blood concentration or ${\mu }_{\mathrm{eff}}$ and compares to its corresponding value with the correction function applied (orange).

3.2.

Verification of heterogeneous phantoms

Once the correction technique was verified, it was applied to a phantom series featuring synthetic blood vessel inclusions with varying inner diameters (${Ø}_s=0.8\mathrm{mm}$, ${Ø}_m=1.9\mathrm{mm}$, and ${Ø}_l=2.3\mathrm{mm}$), and contrasts were compared between corrected and uncorrected images. In Fig. 5(a), spatial frequency domain images recover the effective attenuation coefficient ${\mu }_{\mathrm{eff}}$ in the location of blood vessel (top row). The uncorrected CI (row 2) shows attenuated signal in the same region. The two images are registered, and the correction factor defined in the denominator of Eq. (4) is applied to the CI (third row) using the corresponding intensity values from the SFD image. With each iterative addition of superimposed media, the observable attenuation of the Cherenkov signal diminishes in response to increased signal from “tissue” overlying the attenuating vessel, and the SFDI ${\mu }_{\mathrm{eff}}$ signal attenuates as the primarily absorbing vessel becomes covered by highly scattering media. For the smallest vessel [${Ø}_s=0.8\mathrm{mm}$, Fig. 5(a)], the negative contrast increases from 0.92 to 0.98 in the case for 0 mm [Fig. 5(b)]. When 5-mm depth in media is reached, the negative contrast between the corrected and uncorrected CI differs by only 1%. In the next largest vessel [${Ø}_m=1.6\mathrm{mm}$, Fig. 5(c)], negative contrast increases from 0.76 in the uncorrected image to 0.99 in the corrected image at 0 mm, illustrating the best correction and most dramatic improvement [Fig. 5(d)]. At 5 mm, the negative contrast decreases to a 3% difference between corrected and uncorrected images. In the case of the largest vessel [Fig. 5(e)], the minimum negative contrast achieved for the corrected vessel is 0.9 in both cases where vessel depth = 0 and 5 mm. The negative contrast percent difference increases from 29% at 0 mm, to 2% at 5 mm [Fig. 5(f)].

Fig. 5

In (a) (small vessel case, ${Ø}_s=0.8\mathrm{mm}$), (c) (medium vessel case, ${Ø}_m=1.9\mathrm{mm}$), and (e) (large vessel case, ${Ø}_l=2.3\mathrm{mm}$), the SFD image of each vessel is organized at the top of the phantom’s respective CI both uncorrected (middle) and corrected (bottom). The uncorrected CI more clearly shows the synthetic vessel region through the center of each image, whereas the corrected image shows the diminishing of local differences between the two regions after having applied the pixel-by-pixel correction factor, dictated by Eq. (4). The associated negative contrast ($I_{\mathrm{vessel}}/I_{\mathrm{medium}}$) is listed under each CI. The color bar represents scaling for both modalities, where the SFDI axis and CI axis are shown to its right. In (b), the small vessel (negative contrast values are organized into two plots: corrected and uncorrected, as well as for the case for (d) a medium sized vessel ${Ø}_m=1.9\mathrm{mm}$ and (f) a large vessel ${Ø}_l=2.3\mathrm{mm}$.

3.3.

Verification of Patient Images

The technique was then applied to data acquired from Patient 2. The SFDI effective attenuation map in Fig. 6(a) is masked to include only pixels within a height and angle tolerance acceptable for height correction algorithms. In Fig. 6(b), the uncorrected CI shows the presence of vasculature due to attenuation of Cherenkov light in the hemoglobin content of the blood. The other two most prominent features include the nipple and the surgical scar, both also highly attenuating due to variations in pigment content. In Fig. 6(e), the correction of attenuation due to vasculature is showcased, then magnified in Fig. 6(f). As shown in Figs. 6(c) and 6(f), the vasculature in this region of the CI from a fraction of Patient 2’s treatment appears to be completely corrected for, qualitatively. Further, Fig. 6(d) shows the rendered dose map (VTK) associated with one fraction of treatment from the perspective of the Cherenkov camera, sampled down to $\sim 5\mathrm{mm}$, weighted by a short buildup region, followed by exponential falloff.

Fig. 6

In (a), the effective attenuation optical property map is generated and masked to include pixels at 20% the maximum Cherenkov value or greater, as shown in (b). The uncorrected CI exhibits darker regions due to near-surface vasculature, the nipple, and surgical scar. The attenuation introduced by blood absorption due to subcutaneous blood vessels yields a contrast percent difference of up to 19%, emphasized in (c). In (e), the corrected CI is adjusted, pixel-by-pixel, using the methodology discussed in Eq. (4), and magnified in (f) to again emphasize the region, where the SFDI is best in focus. To illustrate an ideal theoretical case, the 5-mm subsurface dose is read from DICOM format, aligned to the same view angle as of Cherenkov and SFD cameras, rendered in VTK, and exported from the Cherenkov imaging software (d).

An ROI mean was taken at three vessel locations inferior to the nipple [indicated in Fig. 6(c)] and at regions directly adjacent to those in order to determine negative contrast values (Table 1). Because negative contrast does not describe a metric of success, a percent difference was taken between values in corrected and uncorrected images and compared to the patient treatment plan, such that the slight intensity gradient associated with changes in dose deposition between ROIs would be accounted for. It becomes evident from Table 1 that percent differences associated with the corrected image lateral, inferior, and medial vessels (6%, 4%, and 7%, respectively) adhered more closely to the intensity gradient associated with the treatment plan, compared to that of the uncorrected image (22%, 14%, and 10%, respectively). Although the medial case does quantify an improvement, the change is notably less than those illustrated by ROI 1 and 2.

Table 1

The CI contrast between a small ROI inside a region attenuated by the vessel and a nearby region adjacent to the vessel are organized into rows 2 and 3, separated by the corrected and uncorrected values (bold). The first row organizes contrasts associated with the same regions of the coregistered treatment. Percent differences between contrasts for each CI ROI and respective treatment plan ROI are provided in the appropriate box.

3.4.

Dynamic Nature of Tissue Optical Properties

SFDI measurements show that tissue optical properties varied significantly throughout the course of a hyper-fractionated radiotherapy treatment for Patient 1. In Figs. 7(b), 7(d), and 7(f), the axillary and mammary fold regions illustrate where erythema due to irradiation is most evident. The associated color images were reconstructed into RGB channels using the modulated images at each phase and spatial frequency of the wavelengths employed using the NIR setting of the SFDI system (649, 691, 731, and 851 nm) and the demodulation equation. ¹⁴ Because blue wavelengths are neglected, the product is essentially a “false color” image. Figure 7(a) shows the tissue optical property breakdown sampled from the ROIs delineated in white from the mammary fold just inferior to the breast [(c), (e), and (g)]. It was shown that reduced scatter decreased over the course of the treatment by $\sim 13\%$, and absorption increased by $\sim 49\%$ for Patient 1. A similar ROI evaluation of the measured Cherenkov light shows substantial Cherenkov falloff as treatment progresses. Overall, the effective attenuation increased by $\sim 24\%$, and the corresponding Cherenkov emission decreases by 22%. Going forward, it is important to acknowledge the highly dynamic and temporally dependent nature of tissue optical properties throughout ongoing radiation therapy.

Fig. 7

(a) The breakdown of changing tissue optical properties in Patient 1, which are normalized and plotted over the course of one month of treatment (beginning, middle, and end), where SFDI data are taken at days 2, 24, and 32, and Cherenkov at days 1, 21, and 27. The region analyzed is indicated by a red circle in the mammary fold in the bottom row of effective attenuation maps (c), (e), and (g). Absorption in this region increases by 49%, and reduced scatter decreases by 13%, yielding an overall effective attenuation change of $\sim 24\%$. Measured Cherenkov emission, conversely, decreases by $\sim 22\%$. Although these changes are not illustrated in full-color format, the false-color images reconstructed using the NIR wavelengths employed by the reflect RS clearly indicate the changes incited around the mammary fold, as well as those in the axillary fold and the nipple-areola complex (b), (d), and (f).