2.1.

Instrumentation

Three Lucid VivaScope™ reflectance confocal microscopes (Lucid Inc., Henrietta, New York, http://www.lucid-tech.com) were used in these measurements: a production VivaScope 1000 and two different VivaScope laboratory prototypes. Of the microscopes tested, only the VivaScope 1000 was being used in clinical testing. The other systems used were imaging prototypes. The goal of the measurements was to test the accuracy and precision of the algorithm to compare the imaging performance of different microscopes. Once validated, this technique could be used to compare the various instruments used in multicenter trials.

VivaScopes are Food and Drug Administration (FDA)–cleared laser-scanning reflectance confocal microscopes equipped with either a $30\times$ 0.9–numerical aperature (NA) LOMO (St. Petersburg, Russia) or a $30\times$ 0.9-NA ASE Optics (Rochester, New York) water immersion objective. The system details have been described previously.⁵ A schematic of the system is shown in Fig. 1 . An $830\text{-}\mathrm{nm}$ laser illumination beam is scanned by polygon and galvanometer mirrors and relayed into the water immersion microscope objective that focuses the beam into the patient. The acquired image is nominally parallel to the tissue interface. The field of view is $0.5\times 0.375\mathrm{mm}$ . The depth of the image plane is adjusted by moving the microscope objective relative to the tissue surface using a $1\text{-}\mu \mathrm{m}$ resolution micrometer. Light reflected from the tissue is collected by the objective, descanned, and focused onto a five-resolution element pinhole. Light passed through the pinhole illuminates a single-element silicon avalanche photodiode. The detected optical signal is digitized by an $8\text{-}\text{bit}$ frame grabber and displayed on a computer monitor with $640\text{-}\times \text{-}480\text{-}\text{pixel}$ resolution. The target used to measure the lateral resolution was a chrome-on-glass stripe $55\text{-}\mu \mathrm{m}$ wide. An image of the target is shown in Fig. 1.

Fig. 1

Optical and electronics design of a laser reflectance confocal microscope (VivaScope).

2.2.

Pixel Clock Asynchronism and Galvo Locked Scanning Mode

As shown in Fig. 1, a 36-facet polygon provides a “fast scan” of the point illumination across the lateral extent of the sample. A galvo mirror sweeps the “fast scan” line across the sample vertically forming the “slow scan” raster. An auxiliary laser reflects off each polygon facet and strikes a split detector to provide a “start of scan” (SOS or HSYNC) signal for each fast scan line of the image. As each HSYNC signal is asserted, an image acquisition system (National Instruments, Austin, Texas) digitizes the detector signal at approximately $5\mathrm{MHz}$ using an independent asynchronous pixel clock (Fig. 2 ). Given this sampling rate and scan speed, there will be $2\text{to}3\text{pixels}$ covering the optical transition across the edge [Fig. 3c ]. These data points are insufficient to calculate a smooth and accurate line spread function (LSF) and MTF.

Fig. 2

Line scan signal (HSYNC) and pixel clock signal. Note the asynchronism between those two control signals.

Fig. 3

Algorithm illustration: (a) Image of the test target at normal scan mode. (b) Image of the test target at galvo locked mode. (c) Edge response from two different lines sampled by the pixel clock before and after dejittering. Note the offset is mostly reduced after dejittering. Inset shows pixel jitter along the physical edge. (d) An oversampled ESF combined from all dejittered lines. (e) LSF calculated from ESF. (f) One-dimensional MTF calculated from LSF.

As shown in Fig. 2, the HSYNC signal and pixel clock is not synchronized and the time delay between these two clocks varies by row. This asynchronous pixel clock adds a random offset within $1\text{pixel}$ to each line of the image. Figure 3a shows an image of our test target at normal scanning mode (galvo unlocked). A strip with sharp edges can be identified in the center. This strip is a line of reflectance of approximately 48% (chrome-water reflection at $785\mathrm{nm}$ ) on a background of 0.4% (water-glass reflection). Comparing the digitized intensity between successive pixel rows near the physical edge, the pixel jitter can be observed [Fig. 3c inset]. Finding the actual offset of each row and removing such jitter will provide an oversampled edge to calculate the LSF and one-dimensional MTF. At normal scanning mode, it is difficult to find those accurate offsets between rows. A small slant of the physical edge will overwhelm the pixel jitter. Image intensity along the edge is space variant [Fig. 3a] due to the field curvature of the given imaging system. Therefore, running the microscope at galvo locked mode is required. When imaging at galvo locked mode, the galvo (slow scan) is turned off and the imaging laser beam is scanned repeatedly over the same lateral line across the object. The acquired image [Fig. 3b] consists of duplications of the central line of the test target when imaged at galvo unlocked mode [Fig. 3a]. Because a single line is imaged instead of the whole field, the vertical extent of the image is space invariant and the influence by a slanted object edge is avoided. The major difference between lines can be attributed to the scan error by the asynchronous pixel clock.

2.3.

Dejitter Using Fourier Shift Theorem

We use a Fourier method to calculate the offset of each line caused by the asynchronous pixel clock. According to the Fourier shift theorem, a delay in the spatial or time domain corresponds to a linear phase term in the frequency domain 1. Using the first line of the image as a reference $\left(I_{\mathit{ref}}\right)$ , the offset of the $j\text{th}$ line $\left({\Delta }_j\right)$ relative to the reference can be calculated from the phase difference $\left(\mathit{PD}\right)$ in the frequency $\left(f\right)$ domain Eq. 2. A linear fit between $f$ and $\mathit{PD}$ to calculate ${\Delta }_j$ is required to reduce error by noise. The jitter is then removed by shifting each line in the amount of $-{\Delta }_j$ [Fig. 3c]. Because each offset is a subpixel variable, an oversampled edge is derived by accumulating all the lines in a single image after dejitter [Fig. 3d]

2.4.

Calculation of Edge Spread Function, LSF, and MTF

The algorithm to calculate the edge spread function (ESF), LSF, and MTF [Figs. 3e and 3f] is similar to ISO-12233. The oversampled data points [small dots in Fig. 3d] are binned by one-fourth the pixel width and averaged [large dots in Fig. 3d]. A continuous ESF is formed from these $4\times$ oversampled data points. The ESF, which contains resolution and MTF information similar to the PSF, is the imaging system response to the real physical edge on the target. Once the ESF is derived, the LSF can be calculated by making a two-point derivative of the ESF. The one-dimensional MTF is a normalized fast Fourier transform of this LSF. Additionally, the full-width half maximum (FWHM) of the LSF is used as a measure of the lateral resolution. A MATLAB program was developed to measure the MTF and lateral resolution from a single image with a sharp straight edge.

2.5.

Validation

An oversampled edge response, which is key to calculating the lateral resolution and MTF, can also be acquired using a fast oscilloscope (Tektronix, TDS5000B, Richardson, Texas) by directly connecting it to the pinhole detector. At the sampling rate of $4\mathrm{ns}$ , the oscilloscope can oversample the edge 50 times faster than the $5\text{-}\mathrm{MHz}$ pixel clock. Those $50\times$ oversampled data points are further binned down to 4 times the pixel rate, which matches the oversampling rate used in our algorithm under test. Then, the lateral resolution and MTF can be derived the same way from these oversampled data provided by the oscilloscope. We use this alternative method to validate the results calculated from our dejitter algorithm. It is important to note, however, that this oversampled edge response by the oscilloscope is from one line scan and thus from one facet scan of the polygon. Our dejitter algorithm, on the other hand, oversamples by phase-shifting multiple line scans, or equivalently, multiple facet scans.

3.1.

MTF Measurement of Different Imaging Systems

We measured the MTFs of three different models (prototypes) of VivaScope confocal imaging systems (Fig. 4 ). As expected, the MTF of each system has its own signature, even if those systems were built using similar optical and electronic designs. The MTF differences influence the image appearance and may bias the image interpretation. Therefore, it is essential to consider the specific MTF of each system when evaluating images, especially when clinicians try to reach unbiased diagnosis consensus based on images acquired from different systems. To examine repeatability, we performed six measurements on the multiwavelength VivaScope prototype (solid line in Fig. 4) with various time intervals from seconds to minutes. Particularly, the spatial frequency at 10% visibility (contrast) is an important specification of the MTF. Results show the average spatial frequency at 10% visibility is $950\mathrm{lp}∕\mathrm{mm}$ , and the standard deviation is $9.8\mathrm{lp}∕\mathrm{mm}$ . Considering this significant order of magnitude difference, the measurement variation within a system can be neglected. It has to be mentioned here that although the ideal Nyquist frequency of a 0.9-NA optical system is about $1200\mathrm{lp}∕\mathrm{mm}$ , real systems may not achieve that due to insufficient electronic sampling. The lateral Nyquist frequency of tested VivaScope systems is $640\mathrm{lp}∕\mathrm{mm}$ , given the pixel resolution $(640\times 480)$ and field of view $(0.5\times 0.375\mathrm{mm})$ . The frequency we measured at 10% contrast, which is higher than the system Nyquist frequency, is derived from the oversampled edge response and represents the actual optical frequency response of the system.

Fig. 4

MTFs of three confocal imaging systems derived from oversampled edge responses. (a) Multiwavelength VivaScope prototype using a $30\times 0.9\text{-}\mathrm{NA}$ LOMO objective and $785\text{-}\mathrm{nm}$ illumination (solid). (b) VivaScope 1500 prototype located at Lucid, Inc., using a $30\times 0.9\text{-}\mathrm{NA}$ ASE objective and $830\text{-}\mathrm{nm}$ illumination (dash dot). (c) VivaScope 1000 located at Memorial Sloan-Kettering Cancer Center using a $30\times 0.9\text{-}\mathrm{NA}$ LOMO objective and $830\text{-}\mathrm{nm}$ illumination (dash).

3.2.

LSF and Lateral Resolution Measurement

Figure 5 shows the LSF of one tested system ( $a$ in Fig. 4). The lateral resolutions of the three systems, which is measured as the FWHM of the LSF, are 0.88, 0.83, and $0.84\mu \mathrm{m}$ (from $a$ to $c$ as in Fig. 4). The same repeatability test of system $\left(a\right)$ as mentioned in Sec. 3.1 shows the standard deviation of six lateral resolution measurements is $0.037\mu \mathrm{m}$ . The difference between the three systems is also small (less than $0.1\mu \mathrm{m}$ ), which implies the MTF is more specific to characterize an imaging system than resolution. Ideally, for a confocal system with $785\text{-}\mathrm{nm}$ illumination using a 0.9-NA water immersed objective, the lateral resolution is about $0.44\mu \mathrm{m}$ . The actual resolutions are much larger in those three systems because of residual system aberrations in the relay optics and objective lens and the large five-resolution element pinhole used to increase signal-to-noise ratio.^{5, 6}

Fig. 5

LSF of the multiwavelength VivaScope prototype [(a) in Fig. 4].

3.3.

Validation

As mentioned previously, the lateral resolution and MTF can be derived in a similar fashion from an edge response oversampled by the oscilloscope. We use this method to validate the results calculated by our dejitter algorithm. Figure 6 shows good agreement in the shapes of the MTF curves calculated from the two different methods. At 0.9-NA setting, the spatial frequency difference at 10% visibility $(55\mathrm{lp}∕\mathrm{mm})$ is only about 9% of Nyquist frequency of the system $(640\mathrm{lp}∕\mathrm{mm})$ . The lateral resolution difference is less than $0.05\mu \mathrm{m}$ . A small discrepancy exists between MTF curves from two different methods. To test whether this discrepancy is systematic at different instrument configurations, we lowered the objective NA of the testing prototype to 0.35. As expected, the MTF using 0.35-NA objective is significantly narrower than using the 0.9-NA objective. However, the discrepancy between two methods remains consistent and small (Fig. 6). At both NA settings, the MTFs calculated from the edge response given by the oscilloscope are slightly broader than that of the dejitter algorithm, especially at high spatial frequency. In the dejitter algorithm, the oversampled edge response is acquired by accumulating all the lines in a single image. Intensity noise between these lines adds some errors to the offsets found by the Fourier method, which could narrow down the MTF and reduce the lateral resolution. The oscilloscope, however, can oversample an edge from a single image line so the results are less affected by intensity noise. The much higher sampling rate of the oscilloscope $\left(250\mathrm{MHz}\right)$ may also explain the broader MTF at high spatial frequency. However, the limitation of using an oscilloscope is also obvious. Because one acquisition of an edge response by the oscilloscope only covers a single line scanned by one facet, to fully describe the system multiple measurements from different facets are required for an average, which is time-consuming. Users also may not have the access to reconnect the detectors to an oscilloscope. Furthermore, the availability of a fast digital oscilloscope at a clinical trial site can itself be a problem.

Fig. 6

MTFs of the multiwavelength VivaScope prototype at two different NAs (0.9 and 0.35) calculated from edge response oversampled by two different methods: dejitter algorithm [solid (a) for 0.9 NA, solid (b) for 0.35 NA], oscilloscope (dash-dotted line for 0.9 NA, dashed line for 0.35 NA). Because oscilloscope only scans a line for one measurement, the MTF presented here is an average of 10 different measurements to better describe the whole system.