Digitl Cmer Technologies for Scientific Bio-Imging. Prt 4: Signl-to-Noise Rtio nd Imge Comprison of Cmers Yshvinder Shrwl, Solexis Advisors LLC, Austin, TX, USA B I O G R A P H Y Yshvinder Shrwl hs BS in optics from the University of Rochester nd MS nd PhD in opticl sciences from the University of Arizon. He ws cofounder of Opticl Insights LLC where he ws responsile for the development of imging pltforms for fluorescence microscopy, high-throughput screening, nd high-content screening. Following the sle of Opticl Insights, he mnged product mrketing for scientific digitl cmers t Photometrics. Since 2006, Dr Shrwl hs worked s technicl nd usiness consultnt with vrious compnies nd entrepreneuril orgniztions. In 2010, he moved to Austin nd is currently the COO for Xeris Phrmceuticls. A B S T R A C T In this rticle, the lst in four prt series, the signl-to-noise rtio (SNR) comprisons egun in Prt 3 re expnded with exmples of existing cmers in low light nd medium-to-high light scenrios. The opticl system throughput clcultions re comined with collection, conversion, nd noise specifictions from typicl cmers ville on the mrket to generte SNR vlues t different exposure times nd under common user scenrios. Direct quntittive comprisons of imge qulity re mde using SNR grphs nd experimentl imges. K E Y W O R D S light microscopy, digitl cmers, CCD, EMCCD, scientific CMOS, imging, life sciences A U T H O R D E TA I L S Yshvinder Shrwl, Solexis Advisors LLC, 6103 Cherrylwn Circle, Austin, TX 78723, USA Emil: ysh@solexis-dvisors.com I N T R O D U C T I O N The previous three prts of this series covered lrge mount of ground in compring the three most common cmer technologies for io-imging CCD, EMCCD, nd CMOS sensors. They provided n introduction to the different technologies [1], discussed their differences, nd compred their performnces in terms of smpling requirements [2], light throughput t the detector, noise types, nd signl-to-noise rtios (SNR) [3]. The gol throughout the series hs een to utilize relistic opticl prmeters of n imging system in the selection of cmer for prticulr iologicl imging pplictions. In this finl prt of the series, the SNR comprisons egun in generl sense in Prt 3 [3] will e expnded to exmples of existing cmers in low-light nd medium-to-high-light scenrios. The opticl system throughput clcultions will e comined with collection, conversion, nd noise specifictions from typicl cmers ville on the mrket to generte SNR vlues t different exposure times nd under common user scenrios. Direct quntittive comprisons of imge qulity will then e mde using SNR grphs nd experimentl imges. T W O WAY S T O C A L C U L AT E S I G N A L - T O - N O I S E R AT I O S There re two common pproches to determining the signl-to-noise rtio. One involves clculting the signl t the detector sed on known cmer nd opticl system specific- tions nd then dividing y the noise from the signl nd cmer to otin the SNR. The other pproch is more empiricl. It involves mesuring the noise (stndrd devition) nd men signl in ech pixel in cptured imge nd then dividing the two to clculte the SNR t ech pixel. The pixels contining the fetures of interest cn then e verged together to otin n verge SNR. If the correct opticl nd cmer prmeters re used nd the imges crefully cquired nd mesured, the two methods of clculting SNR should rech closely mtching vlues. Clculting the SNR sed on opticl nd cmer specifictions hs een covered in Prts 2 nd 3 [2, 3]. Briefly, dye concentrtion nd moleculr rightness determine the light flux t the smple, which is collected with certin efficiency determined minly y the numericl perture of the collection optics. The light flux hitting ech pixel then depends on the spred of the imge y the ojective mgnifiction nd y the size of ech of the cmer s pixels collecting the light. The electronic signl generted then depends on this per-pixel light flux, the quntum efficiency (QE) of conversion of light to electrons, nd the exposure time during which the light is collected. The noise sources minly photon shot noise, red noise, nd drk current noise cn e determined from the signl, the time, nd either the cmer specifictions or empiricl mesurements. These noise vlues re then dded in qudrture to otin the totl noise for clculting the SNR per pixel. Figure 1: Outline of procedure to empiriclly clculte the verge signl-to-noise rtio (SNR) in n imge from stck of smple imges nd stck of is imges. Microscopy nd Anlysis 26(1):S4-S8 (AP) 2012 S4
E M P I R I C A L C A L C U L AT I O N O F S I G N A L - T O - N O I S E R AT I O As mentioned ove, clculting the verge SNR in n imge cn lso e determined empiriclly using stck of the sme smple imge nd stck of is imges. The process used is outlined in Figure 1 nd in more detil in previous pper [4]. Briefly, the SNR cn e mesured using stck of smple imges nd stck of is imges. First, stck of 10 imges is cquired for cmer t given exposure time. The stndrd devition nd men of ech pixel through the stck is then found. The verge is vlue is next sutrcted from the men smple imge to otin signl t ech pixel. This imge is then divided y the stndrd devition imge, or noise, to otin the SNR t ech pixel. The men nd stndrd devition of ckground re is mesured to determine the threshold, which is set t one ckground stndrd devition ove the ckground men. Finlly, the men vlue of the SNR of ll the pixels tht re ove this ckground threshold is mesured to find n verge SNR for the smple fetures in smple imge. C O M PA R I S O N O F S N R VA L U E S F R O M B O T H A P P R O A C H E S If the red noise, mgnifiction, nd other vlues for given cmer/microscope system re ccurte nd vrious fctors such s system trnsmission re tken into ccount, the SNR vlues clculted from the system prmeters should mtch firly closely the empiricl dt otined from smple imges. To demonstrte this, grphs of SNR versus integrtion time were generted for two cmers with known mesured red noise, drk current, nd other prmeters. This SNR grph is displyed in Figure 2. In this cse two cmers with the sme type of sensor scientific-grde complimentry metl oxide semiconductor (CMOS) were chosen for the comprison. The two cmers, however, differed in pixel size nd, consequently, in smpling requirements. A moderte light level common to mny microscopicl io-imging pplictions ws used over rnge of typicl exposure times for the comprison. Here, deep-cooled scientific CMOS cmer (scmos, Neo from Andor Technology) with 6.5-µm pixels ws compred to nother scientific CMOS cmer (Roler Bolt from QImging), which hs smller 3.63-µm pixels. As descried in Prt 2 of this series [2], Nyquist smpling requires tht the distnces resolvle y the microscope optics e mgnified to t lest twice the size of the cmer sensor s pixels. For exmple, with 60 ojective, the deep-cooled scmos s 6.5-µm pixels cn dequtely Nyquist smple specimen. However, the 3.63-µm pixel scientific CMOS would oversmple. Thus ecuse it hs smller pixels, it cn use lower mgnifiction while still Nyquist smpling. To chieve this lower mgnifiction for the 3.63-µm pixel scientific CMOS while still Nyquist smpling, one of the newer 30 or 40 ojectives with high NA cn e used, or 0.5 coupling lens cn e dded efore the cmer when 60 ojective is used. Both options provide lower mgnifiction of the imge hitting the sensor so tht the imge is much less spred out, nd thus it is condensed down efore projection onto the cmer sensor. This provides two dvntges: 1. Condensing the imge concentrtes more light per pixel to oost signl nd SNR; nd 2. Less spreding out of the imge llows more of the imge to fit on the sensor for lrger effective field of view. Three common exposure times were chosen, s indicted y the dshed lines in Figure 2, nd used to cquire imge stcks t the pproprite light levels of the sme smple using the two cmers. Simultneous imging of the sme smple y the two cmers ws performed y using DulCm DC2 multichnnel system with 50/50 emsplitter to project the imge eqully onto the two cmers. The cquired imges re pictured in Figure 3. The SNR vlues empiriclly clculted from ech imge re lso included in Figure 3. Note tht ll of the imges in Figure 3 nd in other figures in this rticle re zoomed in to mke the fetures pper the sme size for imge qulity comprison, despite differences in the Exposure 50 ms Exposure 100 ms Figure 2: Grph of signl-to-noise rtio (SNR) for 6.5-µm pixel deep-cooled scientific CMOS nd 3.63-µm pixel scientific CMOS cmer t incresing exposure times. Verticl dshed lines indicte smple integrtion times chosen for empiricl comprison in Figure 3. fields of view nd pixel sizes. All imges hve lso een scled so tht the minimum nd mximum intensities displyed re set to the intensities t the sme loctions in ll imges within the sme figure. As cn e seen y compring Figures 2 nd 3, the SNR grph vlues clculted from cmer nd system prmeters mtch the vlues otined from imges cquired under those conditions resonly well. Additionlly, it is noteworthy tht lthough the 3.63-µm pixel scientific CMOS produces imges with overll lower SNRs thn the deep-cooled 6.5 µm pixel scmos, the difference is not extremely lrge. The imge qulity is lso similr etween the two sets. The deep-cooled 6.5-µm pixel scmos does, however, hve two other min dvntges over the 3.63-µm pixel scientific CMOS in tht it hs fster 100 frmes per second short urst mximum full frme rte nd lrger field of view. Additionlly, t lower light levels, the difference etween the two cmers ecomes more significnt. In low light situtions, the low red noise in the deep-cooled 6.5-µm pixel scmos would give it noticele dvntge over the 3.63-µm pixel scientific CMOS s this would e red noise-limited Exposure 200 ms 9.4 13.9 19.6 7.7 11.4 16.4 Figure 3: Comprison of signl-to-noise rtios (SNR) of differently exposed imges of the sme smples of fluorescent owl monkey kidney cells. () The deep-cooled 6.5-µm pixel scientific CMOS. () 3.63-um pixel scientific CMOS. Note tht other issues were discovered for the 6.5-µm pixel scientific CMOS imges which re displyed in Figs 7, 8, nd 9 (See Unexpected Findings). S5
scenrio. It is lso useful from cmer selection viewpoint to compre the effect tht cmers with vriety of pixel sizes, red noises, nd other prmeters hve on SNR under different experimentl scenrios. To this end, SNR grphs of different cmer types with such vrying prmeters cn e produced using light levels commonly encountered in ioimging experiments. These cn e split into two mjor groups covering mjority of iologicl imging pplictions: low light levels, nd moderte light levels. C O M PA R I S O N O F S I G N A L - T O - N O I S E R AT I O I N L O W L I G H T Under low-light conditions, cmer sensors re photon-strved, nd imge qulity is consequently limited in terms of noise, most often y the cmer s red noise. These situtions typiclly encompss fluorescence imging with low signl resulting from low excittion power, low fluorophore density, or short exposure times. These conditions re typiclly required for dynmic live-cell imging, fluorescent protein imging, spinning disk confocl microscopy, superresolution microscopy, nd single molecule fluorescence, mong others. Electron multiplying CCD (EMCCD) cmers hve een developed for this type of imging, s they possess the ility to increse the detected signl orders of mgnitude efore reding out. Since red noise is imprted during the redout process, this multiplictive ility eforehnd mkes their effective red noise orders of mgnitude lower, or less thn hlf n electron. Such low red noise is tremendous dvntge in lowlight scenrios where red noise is the mjor limiting fctor. Scientific CMOS cmers hve recently een developed with red noise vlues s low s round 1 e- when not imging t mximum speed nd depending on how red noise is defined. While this is still somewhere round n order of mgnitude higher thn some EMCCD cmers, it my provide enough sensitivity for some low-light imging. Figure 4 compres the SNR response of two of thesecmer types under low-light conditions nd typicl exposure times: n EMCCD cmer (Evolve 512 from Photometrics) nd deepcooled scientific CMOS cmer (Neo from Andor Technology). It is evident from the lue nd red curves in Figure 4 tht the EMCCD hs much higher SNR vlues thn the 6.5 µm deep-cooled scmos under these conditions. This is result of severl fctors. The first is the EMCCD s low effective red noise, which is pproximtely one order of mgnitude lower thn the scientific CMOS. The second reson is the EMCCD s pproximtely 93% pek QE, which is significntly lrger thn the 6.5 µm deep-cooled scmos s 57% pek QE. The finl reson is the EMCCD s lrger 16 µm-wide pixels, which re little over doule the width (nd qudruple the collection re) of the deep-cooled scmos s 6.5 µm pixels. The pixel size cn e somewht equlized y softwre inning the deep-cooled 6.5-µm pixel scmos. Since it is Figure 4: () SNR versus integrtion time grph compring n EMCCD with deep-cooled 6.5-µm pixel scientific CMOS with nd without softwre inning. () Imges of the sme humn osteosrcom U2 cells tken y the EMCCD (left) nd the deep-cooled 6.5 µm scientific CMOS with nd without softwre inning t 50 ms. Evolve scmos scmos 2x2 inning Figure 5: () SNR response t incresing integrtion times for oth the EM-C2 EMCCD nd the 6.5 µm deep-cooled scientific CMOS t low light levels. () Fluorescent fox lung firolst imges tken t 50 ms with the two cmers: left EM-C2; right CMOS. S6
CMOS sensor, the pixels cnnot e inned onchip, the method of inning employed in CCDs nd EMCCDs. As result, the mximum improvement in SNR cnnot e relized. However, with 2 2 softwre in, the deepcooled 6.5 µm scmos s pixels re effectively closer in size to the EMCCD, nd this gives n pproximtely 2 improvement in the SNR. This grph is plotted s the green line in Figure 4, which rings the response closer to tht of the EMCCD ut still well elow it. The differences in imge qulity mong these three re demonstrted in the imges in Figure 4. While EMCCDs re commonly used for low light imging, not ll re ck-illuminted like the EMCCD tested here. Other EMCCDs, such s the Roler EM-C2, re front-illuminted, which lowers the QE somewht (pek QE round 62%) ut lso lowers the cost to produce. However, these EMCCDs still disply less thn hlf n electron of effective red noise s result of their EM gin. The low light SNR response of the EM-C2 is plotted in Figure 5, long with tht of the deep-cooled 6.5 µm s CMOS. The EM-C2 hs higher SNR performnce for the rnge covered due to the much lower effective red noise nd slightly higher QE. At longer exposure times, however, the 6.5 µm deep-cooled scmos will meet nd egin to overtke the EM-C2 due to the excess noise fctor s effect t higher light. Representtive imges of the two cmers re shown in Figure 5, demonstrting higher SNR qulity imges for the EM-C2 EMCCD. Overll, these clcultions nd imges demonstrte tht for low-light imging, EMCCDs re the sensors of choice due to the oost in SNR given y their electron multiplying cpilities, even for front-illuminted EMCCDs. Figure 6: () SNR plot of the HQ2 CCD nd deepcooled 6.5 µm scientific CMOS t medium-light levels. () Fluorescence imge comprison of rit kidney RK-13 cells tken y the two cmers. Left: HQ2; Right: scmos. Note the dditionl chrcteristics of the scientific CMOS cmer in the Unexpected Findings section nd in Figures 7, 8 nd 9. ^ C O M PA R I S O N O F S I G N A L - T O - N O I S E R AT I O I N M O D E R AT E L I G H T A lrge percentge of iologicl imging pplictions cquire light t moderte light levels. From immunofluorescence nd rightfield imging to GFP nd clcium imging, moderte light levels descrie wide rnge of typicl io-imging situtions. Under these higher photon fluxes, red noise is still importnt ut to lesser degree. This occurs s inherent photon shot noise ecomes more dominnt, equlizing mny sensors to degree, lthough EMCCDs hve dded noise in this higher light regime due to their excess noise fctor. These different noise types re discussed t length in Prt 3 of the series [3]. An exmple of n interline CCD scientific cmer with the uiquitous Sony ICX-285 chip typiclly employed in this rnge is the CoolSNAP HQ2. The 6.5 µm deep-cooled scientific CMOS descried erlier is lso designed for this rnge of light levels nd performs similrly to the HQ2 in terms of SNR in Figure 6. It is pprent tht, lthough the SNR responses of the two curves re close, the deep-cooled 6.5 µm scientific CMOS s response is slightly higher for shorter exposures. However, lthough the red noise is lower for the deep-cooled 6.5-µm pixel scientific CMOS nd the pixel sizes re lmost identicl, inherent photon shot noise Figure 7: Histogrm of deep-cooled 6.5 µm pixel scientific CMOS cmer 16-it imge of fluorescent cell smple. Spcings of 32 intensity units per gp indicte histogrm stretching of 11-it dt to 16 it. egins to dominte t moderte light levels nd gretly equlizes the two cmers. In fct, since the pek QE is little higher for the HQ2 compred to the 6.5 µm deep-cooled scientific CMOS, the HQ2 meets nd overtkes the 6.5 µm deep-cooled scientific CMOS t longer exposures. The imge comprison in Figure 6 shows little noticele difference etween the two cmers t these light levels, s expected from the SNR grph. It should e lso noted tht for low light scenrios, the lower red noise of the deep-cooled scientific CMOS cmer would provide somewht higher SNR response thn the HQ2. U N E X P E C T E D F I N D I N G S In the SNR imge qulity nlyses descried ove, quntifiction issues with the different cmers ppered, minly in the deep-cooled scientific CMOS imges. The first chrcteristic ws in the intensity histogrms of the deepcooled 6.5-µm pixel scientific CMOS imges. This prticulr scmos cmer switches etween two gin mplifiers nd ssocited nlog-to-digitl (A/D) converters to convert detected photoelectrons in ech pixel into n intensity vlue. Ech A/D converter is reported to hve it depth of 11 its, which is sid to e comined on the cmer into 16-it S7
imge. However, nlysis of the histogrm in Figure 7 shows gps etween histogrm points. This occurs over rnge of in smplings nd consistently shows gp of 32, or 2 5, the exct difference etween 11 its nd 16 its. Overll, this highly suggests imge histogrm stretching; indeed, when the cmer is supposed to e delivering 16-it dt, it is ctully only using 11 it dt nd mthemticlly extrpolting. For some imging pplictions, these stretched 11 it imges my e dequte, ut more demnding high-end pplictions my show discrepncies from true 16- it imges. Another issue ffecting imge qulity nd relted to the use of two gin mplifiers ppers in the curve in Figure 8. The figure shows wht is generlly known s men-vrince curve or photon trnsfer curve for the deep-cooled 6.5-µm pixel scientific CMOS. Bsed on photon shot noise, property inherent to ll microscopy light sources, the vrince in the signl should increse linerly with the men signl intensity. Scientific cmers typiclly disply highly liner men-vrince curve over their liner rnge nd hve slope relted to the cmer s gin. However, for the 6.5 µm scientific CMOS cmer, two different gins re used, nd the gin used for ech pixel depends on the intensity level. This is demonstrted in the two stright-line regions with differing slopes (nd ssocited gins) tht cross-over ner the middle of Figure 8. While this hs the enefit of incresing dynmic rnge, hving multiple gins in the men-vrince curve tht vry depending on intensity mkes converting to stndrd quntittive comprison unit, such s the photoelectron, much more difficult. Other issues pper t the low nd high ends of the curve. At the low end, jump occurs in the grph, presumly where switch from unstretched 11 it to stretched 11 it dt occurs. Additionlly, the slope chnges slightly, nd the two lines do not ctully meet. At the high end, insted of quick drop ner the sturtion intensity vlue expected for scientific cmers, there is grdul decrese nd n unusul immense jump to lmost 50 times the vlues oserved in the liner regions (not shown ecuse out of scle) efore dropping to zero (see Figure 8). Furthermore, the cross-over regions etween the two sets of liner rnges show unusul ehvior tht could further ffect quntifiction of imge fetures recorded t these intensities in prticulr. Hving these mny distinct regions in which cmer ehvior vries ccording to verge intensity level cn e prolemtic for quntittive dt nlysis. This ehvior is prticulrly troulesome for imging pplictions where quntittive comprisons need to e mde of intensities expected to vry cross those three different intensity regions either over time or cross the imge. Fluorescent recovering fter photoleching (FRAP) imging is such n exmple, where quntittively monitoring the increse in intensity in drk photoleched spot to determine fctors such s moleculr diffusion rtes would e suspect. It should e noted tht, for the 3.63-µm scientific CMOS cmer nd the scientific CCD nd EMCCD cmers shown here, the menvrince curves hve just single consistently liner region. One finl issue is the use of two different redout directions for two hlves of the chip. This chip rchitecture doules frme rte ut dds n dditionl lyer of vrition to the 6.5 µm scientific CMOS to do so. Figure 9 demonstrtes this vriility etween chip hlves. It mnifests s distinct difference in the verge intensity levels for two hlves of the sensor, even with fltfield distriution of light cross the sensor. Furthermore, this intensity vrition ppers over rnge of light levels. This suggests even further vrition in the gin in the two hlves t oth low nd high intensity levels, significnt concern for quntifiction nd noise/vriility. Once gin, it should e noted tht the 3.63-µm pixel scientific CMOS nd the CCD nd EMCCD cmers used here do not disply this hlf-chip performnce vrition. This unexpected findings section illuminted some interesting findings with respect to the new scientific CMOS deep-cooled cmer nd highlighted some differences etween its current ehvior nd tht which resercher would oserve using the more widely ccepted technologies. Upon considertion of the fct tht this technology is quite new, it is proly not surprising tht such differences exist, s new technologies tend to provide new issues s they go through mturtion period. For exmple, when the world s first EMCCD microscopy cmer ws relesed in 2003 (the Cscde:650), it utilized the Texs Instruments TI TC253 chip; however, this chip which lthough t the time seemed to e the est low light imging CCD ville due to its novel electron multipliction ws quickly superseded y much more cple sensors such s the CCD87 nd CCD97 y e2v technologies. Such mturtion of new technologies would seem to e more norml thn not indeed ll dt on the scientific CMOS presented here were tken with deep-cooled version of the cmer, nd other implementtions my exhiit different ehviors. Figure 8: The photon response curve (men-vrince curve) tken using 6.5-µm pixel deepcooled scientific CMOS cmer. Note the three different liner regions where the cmer responds differently to light nd the chrcteristic ehvior circled etween these regions. Figure 9: Full frme imge from deep-cooled 6.5-µm pixel scientific CMOS cmer with fltfield light source. Employing two different redout directions produces differences in intensities in the two imge hlves. C O N C L U S I O N S This series hs introduced the different sensor types in cmers most often used in iologicl imging, discussed the key differences mong the different types in terms of criticl prmeters such s red noise nd redout modes, nd even discussed the importnce of the vrious noise sources nd their influence on signl-to-noise rtio nd imge qulity. A system of clculting SNR for ech sensor y propgting light throughout n opticl system from smple to detector ws lso proposed nd used to generte grphs of SNR versus exposure time to compre the different sensor types. This ws further expnded upon in this prt of the series to demonstrte how the different cmer types compre under different imging ppliction conditions in terms of oth SNR vlues nd smple imge qulity. As mentioned erlier, the gol of this series ws not to stte tht one sensor technology ws the est performing under ll circumstnces. Rther, this series imed to help provide vriety of fctors to consider, supported y severl dt comprisons of cmers in ppliction scenrios, to ssist in choosing cmer for n ppliction of interest. R E F E R E N C E S 1. Shrwl, Y. Digitl Cmer Technologies for Scientific Bio- Imging. Prt 1: The Sensors. Microscopy nd Anlysis. 25(4):S5-S8 (AM), 2011. 2. Shrwl, Y., Jouert, J., Shrm, D. Digitl Cmer Technologies for Scientific Bio-Imging. Prt 2: Smpling nd Signl. Microscopy nd Anlysis. We Fetures. July 2011:1-5. 3. Jouert, J., Shrwl, Y., Shrm, D. Digitl Cmer Technologies for Scientific Bio-Imging. Prt 3: Noise nd Signl-to-Noise Rtios. Microscopy nd Anlysis. We Fetures. Sept 2011:1-4. 4. Jouert, J., nd Shrm, D. Using CMOS Cmers for Light Microscopy. Microscopy Tody 19(4):22-29, 2011. 2012 John Wiley & Sons, Ltd S8