Seeing the last crystals: second harmonic generation imaging of cryomilled griseofulvin powder

Background and motivation

The paper starts from the solid form of a drug powder. The introduction stresses that the solid form of an organic powder drives its chemical and physical behaviour and that even trace residual crystallinity can significantly affect shelf life. C1 There are two reasons to want to see crystallinity below 1%. First, crystals left behind by incomplete vitrification can "potentially" seed and accelerate crystallization of the amorphous material through heterogeneous nucleation, which can bias any study of crystallization kinetics or mechanism (Fletcher 1963; Carter and Ward 1993; Schmitt et al. 1999). Second, sub-percent detection helps screen preparation, formulation and storage conditions for crystallization control. C3

The standard tools are blunt for this purpose. Powder X-ray diffraction (PXRD), Raman and infrared spectroscopy and differential scanning calorimetry (DSC) all average over a large sample volume, and they give no microscopic information such as crystal size distributions or nucleation kinetics. X-ray diffraction had detected about 0.2% crystallinity, but only with high-energy synchrotron radiation (Nunes et al. 2005), which is not a routine tool. C2·R18

The paper uses milling as its test case. Milling, sometimes under liquid nitrogen (cryomilling) to limit local heating, is used to reduce particle size and raise dissolution rate, and it is known to destroy crystallinity in inorganic and organic solids. The usual picture is that high shear produces a defective, higher-energy crystal that "may subsequently" undergo a crystal-to-glass transformation, but the governing mechanism for organic systems was open. C4 PXRD cannot settle it: loss of Bragg intensity together with a diffuse halo marks disorder, but the halo can come from a true glass (thermodynamic disordering, as in a melt quench) or from crystals broken into nanocrystalline domains that keep their unit cell (kinetic disordering; Bates et al. 2006). In the authors' words, "X-ray amorphous" is not necessarily complete disordering. C5

The method the authors bring is second order nonlinear optical imaging of chiral crystals (SONICC), developed in the Simpson group for detecting protein crystals (Wampler et al. 2008; Kissick et al. 2010) and used by the same authors to detect crystals of active pharmaceutical ingredients (APIs) in transparent amorphous films (Wanapun et al. 2010). C6·R7 Those earlier studies were done in optically clear media. Powders scatter light strongly, and it was not known whether the method would work there. The paper therefore sets out to optimize SONICC for powders, to compare its crystallinity values with PXRD on the same cryomilled griseofulvin samples, and to use the images of individual crystallites to ask how milling destroys crystallinity and how the milled material recrystallizes. A1·A2·A3

Background concepts

Second harmonic generation (SHG). In an intense laser field, a material can convert two photons at a fundamental frequency into one photon at twice the frequency; here 800 nm light in gives 400 nm light out. P3 Because SHG is a coherent process, the contributions of individual molecules add up only when the structure lacks an inversion centre. Most chiral crystals (crystals whose unit cell is not superimposable on its mirror image) therefore generate bulk SHG; the only chiral exception is the cubic point group O. Racemic crystals, with both enantiomers in a centrosymmetric cell, give none, and neither do isotropic media such as liquids, solutions and glasses, where the molecular contributions cancel by destructive interference. C6 For an amorphous drug this is exactly the contrast one wants: the glass is dark and any chiral crystal lights up. Griseofulvin, a single enantiomer, crystallizes in an SHG-active form. M1·R9

Why scattering does not blur the image. SHG grows with the square of the incident intensity, so it is generated only in the tightly focused volume. E4 The authors explain that only "ballistic" (unscattered) light reaching the focus produces signal, so any SHG photon that is detected must come from that focal volume, however many times it is scattered on the way out. Scatter reduces the local intensity, and therefore the signal, but not the spatial resolution; together with the infrared excitation wavelength and the selectivity of SHG, this is why crystals can be located inside a turbid powder that looks opaque in a brightfield image (Figure S-1). C23·R10

Measuring crystallinity by PXRD. The comparison method splits each diffractogram into crystalline peak area and amorphous halo area and takes the crystalline share (unnumbered equation, p.4747):

In plain terms: the degree of crystallinity $X_c$ (in %) is the crystalline area $A_c$ divided by the total area, crystalline plus amorphous $A_a$. The equation assumes that each area is proportional to the mass fraction of its phase, a linear relation that earlier studies had confirmed for APIs and for griseofulvin. E2·C10 For SHG, the authors define the relative crystallinity in words only: the mean SHG intensity over the field of view of a sample, divided by that of 100% crystalline griseofulvin. This holds if crystal orientations in the probed volume are random, which the authors consider reasonable for a powder, and if the detector response stays linear, for which they kept the laser at 20 mW. E1·C8 No polymorph change was seen, so the system could be treated as two phases, crystal and amorphous. R19·C9

One-step versus multistep amorphization. Two pictures of milling-induced crystallinity loss are compared. In a one-step process an intact microcrystal is turned into glass at once (Bates et al. 2006); in a multistep process crystals are broken into ever smaller crystals. The two predict different things for the size distribution of the surviving crystals: the multistep process shifts the distribution toward small crystals as milling continues, while the one-step process changes only the amount of crystallinity, not the shape of the normalized distribution. C26 If every crystallite has the same fixed chance of being amorphized per unit milling time, the crystalline fraction falls exponentially with time; this is first-order kinetics, and the half-life is the time for crystallinity to halve. C19·E5

Nucleation, critical nucleus and induction time. A glass crystallizes by nucleation followed by growth. A new crystal must first form a cluster larger than the critical nucleus; smaller clusters dissolve again. Forming such clusters takes time, which shows as an induction time before the first crystals appear. A domain that is already ordered and larger than the critical nucleus removes this barrier: it can grow at once. This is the idea behind the authors' interpretation of the recrystallization data. C37·C38

Methods in detail

The work has three parts that build on each other: SONICC is first established as a crystallinity assay in powder by comparison with PXRD; the images of single crystallites are then used to test how milling removes crystallinity; finally, DSC and isothermal recrystallization probe what the milled "amorphous" material actually is. The workflow is shown below.

Contribution 1: SHG imaging as a crystallinity assay in a scattering powder

Problem. PXRD stops seeing crystals at a few percent in a laboratory instrument, and the earlier SONICC work was done in clear films. Whether SHG imaging gives quantitative crystallinity in a strongly scattering powder had not been shown. C2·R7

What they did. About 700 mg of crystalline griseofulvin was sealed in the tube of a 6750 freezer mill, cooled for 5 min in liquid nitrogen and milled at 10 Hz in cycles of 5 or 10 min with 1 min of cooling between cycles, the tube kept submerged in liquid nitrogen. P1 Total milling times were 15, 30, 60, 90, 120 and 180 min. P1·A1 The milled powders were warmed to room temperature, kept in a dry nitrogen glovebox and sieved to 50–150 μm, so that particle size would not differ between samples and bias the X-ray or SHG signal. P1 Amorphous reference material was made by holding griseofulvin at its melting temperature for 5 min and quenching it in liquid nitrogen, then lightly grinding and sieving it to the same size range. P2 Each sieved particle is "expected" to be a conglomerate of micro- and nanocrystals and amorphous material. C7

For SONICC, about 2 mg of powder was sandwiched between a slide and a coverslip and scanned with a Spectra-Physics Mai Tai laser (100 fs pulses, 80 MHz) at 800 nm and 20 mW through a 10× objective of numerical aperture 0.3. One frame (1000 averages per line) took 25 s, the focus was stepped in 20 μm increments through the sample, and one voxel measured 4 × 4 × 20 μm³; the measured depth of field and spot size were about 20 μm and about 2 μm. The backward-propagated SHG at 400 nm was isolated with dichroic mirrors and band-pass filters and counted by photomultipliers. P3 The low magnification gave a large field of view, and the low power kept the signal linear from fully crystalline powder downward. C8 PXRD used a Shimadzu XRD-6000 with Cu Kα radiation, scanning 5–35° 2θ at 4°/min in 0.04° steps, with a silicon external standard. P6·M3 The generator setting is printed as "40 kV and 30 mV"; a tube current in mA is presumably meant. X10

Detection limit. The authors estimated the limit from the smallest signal that could be recognized in one voxel: the intensity giving a signal-to-background ratio of 3 in a single 4 × 4 × 20 μm³ voxel, averaged over crystal orientations, divided by the signal of 100% crystalline powder. E3·P5 This is a calculated limit, not a measurement on a sample of known low crystallinity.

Why it works and how it differs. Unlike PXRD, which integrates over the whole irradiated volume, SHG is generated only at the focus and only by crystals, so the amorphous majority contributes no background. The ensemble crystallinity is then the average of many voxels, while each voxel also reports whether a crystal sits there. C6·C23·C40

Contribution 2: single crystallites to test the mechanism of amorphization

Problem. Ensemble methods give the crystalline fraction but not how it is distributed: many partly ordered regions and a few intact crystals can give the same PXRD pattern, and fragmentation and one-step amorphization can give the same loss of crystallinity. C26·C42

What they did. Three-dimensional SHG images were built from z-stacks of crystalline powder and of milled powder (Figure 2). R9 Crystal sizes were obtained with the ImageJ 3-D object counter, which treats adjacent voxels above an intensity threshold (typically 140,000 counts/s) as one crystal; histograms were normalized to the number of crystals in each sample. P7 Because the algorithm cannot separate crystals in neighbouring voxels, it merged crystals at high density: after 60 min of milling the largest "crystal" found was 2.3 × 10⁷ μm³, 25% of the total imaged volume and larger than the starting crystals. Counting was therefore limited to 90, 120 and 180 min, when crystals were sparse. R11·C24

Why it works. Comparing the normalized size distribution at several milling times tests the two mechanisms directly: a constant distribution while crystallinity falls favours the one-step process. C26 Note by the AI writer (not discussed in the paper): the size bins in Figure 3 are multiples of 320 μm³, which is exactly the volume of one voxel, so the smallest bin is a single voxel and the "size" of a crystal is the number of voxels above the intensity threshold. The test can only see crystals of one voxel or more (8–18 μm equivalent diameter, as the authors state), and a bright crystal can occupy more voxels than a dim one of the same size. X12·R12

Contribution 3: energetics and recrystallization of the milled glass

Problem. If the milled powder is X-ray amorphous, is it a glass or a mass of nanocrystals, and does it behave like a glass made from the melt? C5

What they did. DSC (TA Q2000, 10 °C/min from −50 °C to past melting, 2–4 mg in a pinholed Tzero pan, nitrogen purge) was run on the 3-h milled powder, and the recrystallization exotherm was integrated. P9 The measured enthalpy was compared with the enthalpy expected if a fully amorphous sample crystallized, computed from literature values of the heat of fusion and the heat capacity change at $T_g$ (Zhou et al. 2008) with the SI equation:

In plain terms: the heat released when amorphous material crystallizes at a temperature $T$ below the melting point $T_m$ equals the heat of fusion $\Delta H_m$ corrected for the difference in heat capacity $C_p$ between the amorphous and crystalline states over the range from $T_m$ to $T$. The paper reports the result, −53.76 J/g, but not the inputs or the temperature $T$ used. E7·M1·X9 For recrystallization, 3-h milled and melt-quenched powders were held at 50 °C and imaged by SONICC for up to 400 min, and the integrated SHG was plotted against time. P8

Why it works. A crystal-to-glass transformation stores the full excess enthalpy of the amorphous state, so it should all come back on recrystallization; a nanocrystalline product would return only about 40% of the heat of fusion (Fecht 1995). C30·R14 The recrystallization images add what DSC cannot: whether crystals appear at separate nucleation sites after an induction time, or everywhere at once. C35

Experimental analysis

Crystallinity by SHG and by PXRD

The SHG depth profiles in Figure 1a show the signal peaking where the focus sits in the powder layer and shrinking with milling: relative to crystalline powder, the peak fell to about 0.37 after 15 min, about 0.19 after 30 min and about 0.075 after 60 min; after 120 and 180 min it was only about 13.9 × 10⁻⁴ and about 4.6 × 10⁻⁴ (inset of Figure 1a). R1 PXRD (Figure 1b) showed Bragg peaks at 15, 30 and, weakly, 60 min, and only a halo from 90 min onward; after 180 min the pattern could not be told apart from that of melt-quenched glass. The Bragg peaks did not broaden with milling time, which the authors note because broadening would indicate nanostructure formation. R2·R3·C12

Table A. Crystallinity of cryomilled griseofulvin by SONICC and PXRD: values printed in the text, and values read from the log axis of Figure 1c (approximate; converted from log10, see below). R4·R2

The two methods agree at 15 and 30 min, which is the basis for the abstract's statement that SONICC "compared favorably" with PXRD over the PXRD range. C13·R4 At 60 min they differ, 6% by SONICC against about 12% by PXRD; the authors call the difference significant and say it "may arise" from the high halo background in PXRD combined with small Bragg peaks, which degrade the signal-to-noise of the PXRD estimate. R4·C15 After 90, 120 and 180 min, when the powder was X-ray amorphous, SONICC still quantified 0.7, 0.2 and 0.04% crystallinity "with high signal-to-noise". R4·C14 The 120-min value is printed as 0.2% but plotted at about 0.15%, a difference that rounding to one significant figure would explain. X3

The paper describes the extension of the range in three ways: a "three decade" improvement in linear dynamic range (abstract), "4 decades of dynamic range" (results) and measurement "over 3 decades" down to 0.04% (conclusions). The measured values span 100% to 0.04%, which is about 3.4 decades. C14·X2

The estimated detection limit is 4 × 10⁻⁴ %, that is 4 parts per million crystallinity. R6 It is far higher than the about 10⁻⁹ % (about 10 parts per trillion) reported earlier for uniform amorphous griseofulvin films, a difference the authors attribute "largely" to a weak but nonzero background from surface SHG in the powder. R7·C17 They add that the limit can be improved considerably by raising the laser power, because the signal depends on the square of the incident intensity; the power was kept low to keep the high-crystallinity samples in the linear range. C16·E4

The kinetics of crystallinity loss

In Figure 1c the logarithm of the relative crystallinity falls on a straight line against milling time, with a slope of −0.020 ± 0.002 min⁻¹ (8% RSD) and a half-life of 15 ± 1 min. R5 The authors read this as first-order kinetics: the rate of disordering is proportional to the remaining crystalline material, which "suggests that each crystallite has a finite statistical chance" of being rendered amorphous per unit milling time. C18·C19 Extending the line beyond the measured range "suggests" that about 5 h of milling would be needed to reach 1 ppm residual crystallinity. R8·C20

Note by the AI writer (not discussed in the paper): the axis label in Figure 1c says "log" without a base. The plotted values match the printed percentages only for base-10 logarithms (0.7% plots at about −2.15, 0.04% at about −3.4), and the half-life follows from the slope only in base 10: log10(2)/0.020 = 15.05 min, while ln(2)/0.020 would give 34.7 min. The printed "rate constant" is therefore the slope of a base-10 plot; the corresponding natural-log first-order constant is about 0.046 min⁻¹. The same arithmetic reproduces the 5-h extrapolation (6 decades at 0.020 decades per minute is 300 min). X1·E5·R8 The fit is also not perfect: the SONICC points at 90 and 120 min lie about 0.35–0.4 log units below the line and the 180-min point about 0.3 above it, so the last point already departs from the line that the 5-h estimate extends. R5·X5

Where the residual crystallinity sits, and the size distribution

Figure 2 compares 3-D images of crystalline powder and of a milled powder of 0.2% crystallinity. The crystalline powder gives SHG across the whole field; the milled powder gives SHG only at isolated spots, although powder fills the whole field. The depth profile through one such crystallite peaks at about 117 × 10⁴ counts/s. R9 The authors conclude that the residual crystallinity after exhaustive milling comes mainly from a small number of well-localized crystalline domains rather than from many weakly SHG-active centres, and that most of the solid becomes "SHG amorphous" ("suggest"). C21 They also stress that SEM and AFM, which give high-resolution particle images, see only surfaces, whereas SONICC images the bulk of a powder. C22

Table B. Normalized crystal size distribution (fraction of counted crystals per bin) after 90, 120 and 180 min of cryomilling, read from Figure 3 (approximate). R12

Between 90 and 180 min the crystallinity fell about 16-fold, yet the distributions in Figure 3 look alike: about 0.6 of the counted crystals sit in the smallest bin at every time, and the log-frequency inset falls roughly linearly with volume for all three. R12·R4 The authors find "no statistically significant change" and call this "consistent with the one-step process as the dominant mechanism" for crystals of 320–3200 μm³ (8–18 μm diameter if spherical); a multistep process would have shifted the distribution toward small crystals. C27·C26 At the microscopic level, they describe the picture as microcrystals having a finite probability of meeting local shear high enough to deform the lattice and turn amorphous, which would also give the first-order kinetics of Figure 1c. C28 They further infer that the crystals had reached their "fracture limit" before 90 min, since the distribution no longer changed after that time. C25 The diameter range checks out: a sphere of 320 μm³ has a diameter of 8.5 μm and one of 3200 μm³ a diameter of 18.3 μm. E6

Note by the AI writer (not discussed in the paper): no statistical test, crystal count, replicate image stack or error bar is given for Figure 3, and individual bins differ between milling times (for example about 0.57 to 0.63 in the smallest bin, about 0.006 to 0.023 at 1920 μm³). The fracture-limit statement rests on distributions measured only from 90 min onward. X4·X13

DSC: a glass, not nanocrystals

The DSC trace of the 3-h milled powder (Figure 4c) shows no glass transition step, a two-part exotherm (a broad peak near 90 °C and a sharp one near 110 °C, read from the figure) and melting near 217 °C. R13 The authors explain the missing $T_g$ by the exotherm occurring in the same temperature range, as reported earlier for cryomilled griseofulvin, and, following that earlier study (Trasi et al. 2010), describe the bimodal exotherm as surface crystallization at lower temperature followed by bulk crystallization at higher temperature. C29 The recrystallization enthalpy was −52 J/g, against −53.76 J/g expected for a fully amorphous sample. Given the uncertainties of DSC the authors find the two "not significantly different": close to 100% of the excess enthalpy introduced by milling is recovered, "confirming" that any residual crystallinity is a small mass fraction. R13·M1·C31 Because a nanocrystalline product would return only about 40% of the heat of fusion, the calorimetric data "strongly support a crystal-to-glass transformation" on cryomilling. R14·C32 The measured value is a single number without stated uncertainty, and the inputs of the theoretical value are not printed. X9

Recrystallization at 50 °C

Table C. Integrated SHG (× 10⁵ counts/s) during isothermal recrystallization at 50 °C, read from the curves of Figure 4b (approximate). R16

The images in Figure 4a show two different processes. In the milled powder, dim SHG spreads across the whole field within 20 min, bright regions cover much of it by 60 min, and most of the field is bright by 120 min; the authors describe the growth of microcrystals as nearly uniform, with no individual nucleation sites visible. In the melt-quenched powder the field stays dark at first and separate crystals appear, a few at 120 min and more at 180, 240 and 400 min; the authors describe stochastic nucleation with an induction time of 20 min followed by growth of the nucleation centres. R15 The overall crystallization rate of the milled powder is stated as 40-fold faster. R17·C35

The authors' interpretation is cautious. Residual crystals "would be expected" to change crystallization kinetics, acting as sources for secondary nucleation of the same polymorph or heterogeneous nucleation of another (Crowley and Zografi 2002). C33 But SONICC cannot distinguish a true glass from material containing crystalline domains much smaller than about 90 nm, the single-crystal detection limit. C34·R7 The lack of an induction time on the SONICC time scale, the absence of visible nucleation sites and the fast crystallization "suggest that residual crystalline domains remain that were larger than the critical size for nucleation, but below the single-crystal detection limits of SONICC". C37 These domains are "numerous", go "over and beyond" the residual crystallinity seen in Figure 2, are absent from melt-quenched powder, and by the calorimetric data make up a low mass fraction. C36 By removing the nucleation barrier, they make the overall crystallization rate depend only on growth. C38

Note by the AI writer (not discussed in the paper): the paper does not say how the 40-fold factor was obtained, and the melt-quenched curve in Figure 4b carries the label "40×", which could mean either that the curve was scaled up 40-fold for display or that it marks the rate ratio. The 20-min induction time of the melt-quenched powder cannot be checked in Figure 4a, whose melt-quenched frames jump from 0 to 120 min, and the cryomilled curve in Figure 4b itself rises slowly for the first 10–20 min. X6·X7·R16

Discussion

The chain of argument. The authors connect their observations as follows. C19·C27·C32·C37

Steps 1 and 3 rest on direct measurements (Figure 1c and DSC). Step 2 rests on the visual similarity of three size distributions, and the branch for the cryomilled glass rests on kinetics alone: the proposed ordered domains are, by construction, below what SONICC can see. R5·R13·R12·X8

Where the approach applies. The authors list the method's strengths in the conclusions. Bulk methods such as PXRD, solid-state NMR and Raman spectroscopy integrate over large probe volumes, whereas crystal formation begins in highly localized regions; SONICC probes volumes "commensurate" with those of early crystallization and acquires images quickly, which they contrast with Raman microspectroscopy and solid-state NMR, where a single high-resolution image or spectrum can take minutes to hours. C40 The conclusions mention frame rates "of a few seconds" in these studies; the methods section gives 25 s per frame for the data reported here. X11·P3 The authors also state two limits: the analyte must crystallize in a chiral unit cell, and SONICC as developed here cannot easily tell the analyte's crystals from other chiral crystalline components in the probe volume, which "may complicate" analysis of multicomponent powders and tablets. C41

What it suggests for formulation. Note by the AI writer (not discussed in the paper): two practical points follow from the data. First, "X-ray amorphous" is a weak specification. After 180 min of cryomilling the powder still held about 0.04% detectable crystals, and the milled material recrystallized far faster at 50 °C than the melt-quenched glass of the same drug. R4·R17 A milled amorphous drug and a melt-quenched one should not be assumed to have the same physical stability. Second, for chiral APIs SHG imaging offers a way to check amorphous material for residual seeds at levels PXRD cannot reach, before stability studies are interpreted. Both points come from one neat compound; the paper contains no polymer and no dispersion. A1·A2

Limitations

Limitations stated by the authors. The counting algorithm cannot separate crystals in adjacent voxels and overestimates sizes at high crystal density, so size distributions were reported only from 90 min of milling onward. C24 SONICC cannot distinguish truly amorphous material from material containing crystalline domains much smaller than about 90 nm. C34 The detection limit in powder is higher than in thin films because of surface SHG background, and the laser power was kept low to keep high crystallinities in the linear range. C17·C16 The method needs a chiral crystalline analyte and cannot easily discriminate it from other chiral crystals in mixtures or tablets. C41

The AI writer's own judgement (1): what is measured and what is inferred. The strongest results are direct: SONICC and PXRD agree at high crystallinity, SONICC measures to 0.04%, milled powder shows isolated crystallites, and the two amorphous materials recrystallize very differently. R4·R9·R15 The mechanisms are inferences. The one-step amorphization rests on three unreplicated size distributions with no statistics, resolved only down to one voxel, so changes among crystals smaller than about 8 μm cannot be seen. X4·X12 The ordered sub-detection domains are not observed by any measurement in the paper, and the 40-fold rate factor is not explained. X8·X6 The 4-ppm detection limit is a calculation from one voxel, not a demonstrated measurement; the lowest measured sample is 0.04%. R6·R4

The AI writer's own judgement (2): numbers to read with care. The "rate constant" −0.020 min⁻¹ is a base-10 slope (natural-log constant about 0.046 min⁻¹), and the stated ± 0.002 is 10% of the value although 8% RSD is printed. X1 The points depart from the first-order line in a systematic pattern, so the 5-h extrapolation is uncertain. X5 The dynamic range is stated as three and as four decades. X2 The DSC comparison rests on one measurement and on literature inputs that are not printed, and the printed SI equation would give a positive enthalpy for a positive heat of fusion while both reported values are negative, so the sign convention is not stated. X9 Only one compound, one mill, one sieve fraction and one recrystallization temperature were studied. A1·A2

Conclusion

The paper shows that SHG imaging works as a crystallinity assay inside strongly scattering organic powder. For cryomilled griseofulvin it tracked PXRD while PXRD could still see Bragg peaks and continued to quantify residual crystallinity to 0.04%, with an estimated detection limit of 4 ppm. R4·R6·C13·C14 Crystallinity fell first order with milling time, with a half-life of 15 ± 1 min. R5·C18 Single-crystal images showed that the residual crystallinity sits in a few isolated crystallites whose size distribution did not visibly change between 90 and 180 min, which the authors read as "consistent with" whole microcrystals turning amorphous in one step. R9·R12·C27 DSC recovered −52 J/g against −53.76 J/g expected, supporting a crystal-to-glass transformation rather than nanocrystals. R13·M1·C32 The milled glass nevertheless recrystallized about 40 times faster than melt-quenched glass at 50 °C, without visible nucleation sites, and the authors propose that the disordered state "appears to contain" ordered domains larger than the critical nucleus but below SONICC detection. R17·C35·C37