32 September 2019 1527-3342/19©2019IEEE Ex Vivo Breast Tumor Identification T eraher tz waves cover pho- ton energies that are orders of magnitude smaller (0.4– 40 meV) than the visual spec- trum. Therefore, they provide additional information on intrinsic con- densed-matter properties, making them at- tractive for imaging applications in the life sciences [1], [2]. Furthermore, they do not have an ionizing effect and are considered biologically innocuous. The ever-present water in organic matter strongly absorbs terahertz waves, and subtle changes in the water concentration can be indicative of disease [3]. However, the waves’ long wave- length ( ) 3 30 mm m n - severely limits their lateral resolution and creates challenges for high-resolution imaging of biological tissue on the cellular level, e.g., for tumor margin identification during cancer surgery. Currently, tumor margins are exam- ined by a pathologist after surgery, a process that can take several days and, in the case of a positive result, force the Digital Object Identifier 10.1109/MMM.2019.2922119 Date of publication: 9 August 2019 Ullrich R. Pfeiffer (ullrich@ieee.org), Philipp Hillger (hillger@uni-wuppertal.de), Ritesh Jain (rjain@uni-wuppertal.de), Janusz Grzyb (grzyb@uni-wuppertal.de), and Thomas Bücher (buecher@uni-wuppertal.de) are with the Institute for High-Frequency and Communication Technology, University of Wuppertal, Germany. Quentin Cassar (quentin.cassar@u-bordeaux.fr), Jean-Paul Guillet (jean-paul.guillet@ u-bordeaux.fr), and Patrick Mounaix (patrick.mounaix@u-bordeaux.fr) are with the Laboratoire de l’Intégration du Matériau au Système, Centre National de la Recherche Scientifique, University of Bordeaux, France. Gaëtan MacGrogan (g.ma cgrogan@bordeaux.unicancer.fr) is with the Institut Bergonié, Bordeaux, France. Thomas Zimmer (thomas.zimmer@ims-bordeaux.fr) is with University of Bordeaux, France. September 2019 33 Ullrich R. Pfeiffer, Philipp Hillger, Ritesh Jain, Janusz Grzyb, Thomas Bücher, Quentin Cassar, Gaëtan MacGrogan, Jean-Paul Guillet, Patrick Mounaix, and Thomas Zimmer ©ISTOCKPHOTO.COM/DEM10 patient to undergo a second surgery to remove malignant tissue missed the first time. Thus, the ability to examine tumor margins for can- cer cells that extend to the edges (positive margins) during surgery would increase the chance that all cancerous tissue is removed during the first surgery while sparing the healthiest tissue. In the past, basic terahertz re- search has centered on the following fundamental question: Is physics able to provide a robust tool in the terahertz range to accurately dif- ferentiate between different tis- sue types? This question has been addressed by terahertz reflection spectroscopy, which uses reflectiv- ity measurements at 0.2–2 THz to determine the refractive index and absorption coefficient of liquids and biological tissues in the far field. The complex optical properties of both normal and cancerous tissues have been measured by a number of studies in the past. Studies on breast tumors indicate that pulsed terahertz systems are able to distinguish between healthy and diseased tissue [4], [5]. This positive result has led to the development by TeraView Ltd. (Cambridge, United Kingdom) [6] of an intraoperative terahertz imaging system using a portable scanning probe for medical applications. All these promising results suggest that diseased tissue can be identified by permittivity imaging in the terahertz frequency range. However, the broad utiliza- tion of spectroscopic terahertz methods has been held back by the lack of low-cost and compact sensing sys- tems and the diffraction limitation of terahertz waves to resolve tumor margins with the required resolution at the cellular level. Figure 1 shows a conceptual view of the diffraction limit that applies in far-field tera- hertz imagers as compared with a terahertz near-field (superresolution) imager. Because of diffraction, far- field imagers can resolve structural details only on the level of millimeters, whereas near-field sensors can resolve cells on the level of micrometers. This article (an addition to the August 2019 special issue “Wireless Sensors for Biomedical Applications”) examines advances in the development of compact sili- con-based terahertz subwavelength imagers for life sci- ences applications, especially those designed for tumor margin identification in breast-conserving surgeries. For comprehensive surveys of silicon-based far-field imaging devices and components, readers are referred to other review articles, e.g., [7] and [8]. Interaction of Terahertz Waves with Human Tissue Studies on the absorption and refractive index of bio- logical materials in the terahertz region were initiated more than 40 years ago [9]. Several research groups around the world have scrutinized excised tissues for intrinsic contrast at terahertz frequencies to provide useful and unique features for discriminating cancer- ous tumors from healthy tissues. From a spectroscopy point of view, the interaction of biomolecular mate- rial in the terahertz bandwidth is complex because of important collective modes of proteins and polar liq- uids (such as water) that result in high absorption at terahertz frequencies. This limits the depth penetra- tion of terahertz beams, which stimulates the study of the differential responses among various substances in reflection mode. Many works also underline the key role played by water molecules and hydrogen bonds within proteins [10]. For tissue characterization in the terahertz band, one of the first applications on ex vivo human tissue involved imaging excised basal cell carcinoma speci- mens [11]. In vivo work has focused on the skin and accessible external surfaces of the body to measure hydration and tumors. A list of the optical properties of tissue (including blood constituents) for frequen- cies of 0.5–1.5 THz was compiled by the University of Leeds, United Kingdom [12], using a pulsed time- domain system. ©ISTOCKPHOTO.COM/DEM10 34 September 2019 Our work focuses on breast cancer, which is one of the most common diseases among women and is prov- ing to be particularly invasive. As a result of technologi- cal advances in early breast cancer detection, segmental mastectomy or breast conservation surgery is increas- ingly common, limiting the rate of total breast removal. However, the precision with which tumor margins are delineated remains weak, and this occasionally necessi- tates a second surgery to assess whether all of the breast cancer was removed. (See “Tissue Margin Assessment in Breast Conservation Surgery.”) Permittivity-Based Terahertz Imaging for Tissue Contrast In numerous published works, we find investigations of the properties of several types of healthy organ tis- sues, including liver, kidney, heart muscle, leg muscle, pancreas, and abdominal fat tissues, using terahertz pulsed imaging based on the frequency-dependent refractive index and the absorption coefficient of the tissues. In [4], the authors extracted the refractive indices and absorption coefficients for freshly excised healthy and cancerous breast tissues over the frequency range of 0.2–2 THz (the typical usable bandwidth of photonic time-domain equipment). It was found that breast tumor tissue possesses a higher refractive index than both fat and fibrous tissues over this frequency band. Our own experiments with far-field terahertz pulsed imaging (as described subsequently) have shown consistent results. As can be seen in Figure 2, the measured absorp- tion coefficient and refractive index over the frequency range 0.2–0.8 THz can act to differentiate among benign, malignant, and adipose breast tissues. All of the results are the mean values of 10 samples, and the error bars represent a 95% confidence level. However, the imaging techniques used for breast tumors in the previously mentioned studies are limited, not only by measurement factors but also by the selection of imag- ing features to generate the terahertz images. To be more specific, it is not yet known which features offer the best contrast, nor is there an official procedure for feature selection. The refractive index and, consequently, the dielec- tric permittivity offer a promising contrasting mech- anism to distinguish tumorous from healthy tissue. One reason for this could be an increase in the vas- culature associated with faster cell division and higher cell densities in the tumor. These important physiologic changes generally lead to an increase in the water content—to which terahertz imaging is particularly sensitive—and a decrease in the lipid concentration compared with healthy tissue [5]. Note that the spectral data show a smooth variation of the refractive index over frequency, without any reso- nance peaks. Therefore, even relative permittivity characterization at a single frequency should offer a useful contrast for tumor identification and mar- gin assessment. Far-Field Ex Vivo Sample Testing This section summarizes the results obtained by far-field spectroscopic terahertz imaging on freshly excised tis- sue samples, as seen in Figure 3. Here, freshly excised breast tissues were analyzed in reflection mode using a terahertz pulsed imaging system in the 300–600-GHz frequency band to match the latest available frequen- cies with silicon-germanium (SiGe) BiCMOS technol- ogy [19]. Full details are available in [5]. The human tissue analysis and measurements were performed in keeping with the fundamental ethical principles stipulated in the Helsinki Declaration and its later revisions [20]. The written approval of each patient undergoing an excision procedure was obtained before- hand. The measurements were performed inside the hospital on tissues that had not been submitted to any chemical treatments. A commercially available TeraPulse 4000 (TeraView Ltd., Cambridge, United Kingdom), with a modified reflection geometry setup as shown in Figure 3(a), was used in this study. The emitted submillimeter pulses were focused on a biological sample sandwiched between two 2-mm-thick C-cut sapphire substrates mounted on a motor stage to perform reflection tera- hertz imaging. The step sizes in the x and y directions were between 100 and , 500 m n while the far-field lim- ited spatial resolution at the sample was approximately 1 mm at 300 GHz. Coherent photoconductive detec- tion of the reflected pulses was performed using a < ~ 2 Features < ~ 85 Not Resolved Conventional Resolution Limit Features Not Resolved Material Under Test Image Data x 2 Evanescent Field Superresolution Beyond the Diffraction Barrier Figure 1. A conceptual view of a room-temperature superresolution solid-state imaging device, including all components, such as an illumination source, an evanescent- field surface sensor, and a high-responsivity terahertz power detector on a single chip. (From [42].) September 2019 35 photoconductive antenna similar to the one employed for emission. A maximum of 30 min elapsed between the excision procedure in the operating room and the start of the measurement, ensuring no degradation of fresh-state tissues. The double Debye model parameters can be ex- tracted to study and quantify the dielectric response of human tissue at terahertz frequencies [21]. This model targets the mechanism of interaction between terahertz radiation and water molecules and is a com- mon technique applied in spectroscopic terahertz imaging. Our investigations demonstrated that the parameters of the double Debye model can show sig- nificant differences between normal and diseased tis- sue [5]. The measured results are shown in Figure 2. Frequency amplitude images at 500 GHz are also shown in Figure 3(b). These images show a substantial differ- ence between healthy breast tissue and carcinoma in the BiCMOS-compatible frequency range. Tumor areas determined on the images are in good agree- ment with histology slides. Principal component analysis (PCA) was also per- formed over the fast Fourier transform (FFT) ampli- tude image data set in the 300–600-GHz band for each pixel. PCA is a statistical analysis technique used to describe the data and visualize similarities or group- ings by converting data into a set of linearly uncor- related or orthogonal principal components [22]. The first two principal component images of breast tissue are depicted in Figure 4. The first image [Figure 4(b)] exhibits interesting features that can be interpreted as the sample biological structure. Adipose and fibrous Tissue Margin Assessment in Breast Conservation Surgery In contemporary optical histopathology, the standard process for margin assessment is to prepare slices of the excised tumor tissue for evaluation by the pathologist. After optical examination, the pathologist can categorize the tissue margin. For invasive breast carcinoma, a positive margin indicates that the cancer is in contact with the edge of the surgical specimen [13]; for in situ ductal breast carcinoma, a positive margin means that cancerous ducts are within 2 mm of the edge of the surgical specimen [14]. With conventional lumpectomy procedures, 20–40% of excisions are found to have a positive margin [15], [16]. Furthermore, when a new recurrence does arise, 75–90% of cases involve cancerous tissue at the site of the primary surgery. This results in a second surgical procedure to minimize recurrence, entailing additional cost and risking an increased morbidity rate. To date, many techniques have been developed to address the problem of positive margins in breast conservation surgery [17]. The most usual technique for controlling the shape and position of a tumor prior to the surgery is mammography, an imaging technique using X-rays. However, there are some restrictions when trying to record the full extent of a tumor with mammography alone, especially in younger women. Magnetic resonance imaging is another common method for detecting the extent of a tumor prior to surgery, using magnetic coils and a contrast agent, yet this method alone does not demonstrate an improvement in the overall occurrence of positive margins in excised tissue and recurrence of cancer following surgery. In attempts to minimize the occurrence of positive margins, a number of intraoperative techniques are already in use, such as wire-guided marking of the tumor edge, ultrasound and RF detection, and cryoprobe methods to freeze the tumor bulk for better localization of the tumor being excised. Tumor localization alone has not been found to fully address the problem of positive tumor margins. In addition to localization, other techniques attempt to provide assessments of the tumor margin during surgery. Intraoperative specimen radiography evaluates the excised specimen using portable X-ray mammography, but the technique lacks the specificity to be reliable as a standalone method and requires a radiologist during the surgery. For pathology assessment in the operating room, frozen section analysis (FSA) or touch prep cytology is used to perform rapid frozen pathology or pathology of the margin’s surface cells, respectively. While FSA is relatively cheap overall, it increases the length of the operation, necessitates the services of a pathologist close to the operating room, and can require large amounts of tissue, which could interfere with standard pathology diagnosis and tumor staging. Additionally, frozen artifacts in the tissue can make interpretation difficult. If it were possible to detect positive margins in fresh-state tissue in the operating room, other methods, such as RF ablation and local radiation therapy, which have proven successful in treating any detected cancerous tissue remaining in the lumpectomy cavity, could be employed. Consequently, there is a clear need for an operating room device that can accurately define breast tumor margins during breast-conserving surgeries [18] in a simple, fast, and inexpensive manner. 36 September 2019 regions, distinguishable on the visible image, are well defined. The second component shows less contrast in the tissue than the first one. However, in the second image [Figure 4(c)], the borders between the adipose and the fibers of the tissue are well delimited. These two observations on both the first and second compo- nents are interesting because the first component could be used to numerically segment within the sample while the second could serve as an indicator of the tis- sue borders. Clearly, one interoperative challenge is multi- modal spectral histopathology for accurately detecting tumors on the surface of excised breast-conserva- tion surgery specimens in time scales compatible with intraoperative use. New techniques and data analysis algorithms must be developed to obtain objective diag- noses of varied, large specimens free from user vari- ability. For instance, there is a distinct need for a fast technique to accurately identify tumors smaller than 1 mm 2 on large tissue surfaces within 30 min after the surgical excision because current techniques for intraoperative assessment of tumor margins are insuf- ficient in accuracy and resolution to reliably detect small tumors. While a far-field setup, such as the one described previously, fails to address these practical aspects, a near-field imager can precisely localize tis- sue margins and provide insights into the terahertz properties at the cellular level. 1 1.2 1.4 1.6 1.8 2 2.2 2.4 2.6 2.8 3 Refractive Index Frequency (THz) (a) 0 50 100 150 200 250 Absorption Coefficient (1/cm) Benign Malignant Adipose 0.2 0.3 0.4 0.5 0.6 0.7 0.8 Frequency (THz) (b) 0.2 0.3 0.4 0.5 0.6 0.7 0.8 Figure 2. The (a) measured absorption coefficient and (b) refractive index of benign, malignant, and adipose tissue over the frequency range 0.2–0.8 THz. Terahertz Emitter PM1 Terahertz In Terahertz Out 50 mm PM4 PM2 PM3 Terahertz Detector Sample ∆ t Femtosecond Laser Femtosecond Laser Normal Tissue Cancer (a) (b) (c) (i) (ii) (iii) (i) (ii) (iii) Figure 3. (a) The measurement setup. (b) and (c) A comparison of four tissue samples showing (i) histopathological, (ii) visual, and (iii) terahertz images at 500 GHz. Each section contains a healthy region and a malignant one (the dashed area on the histology slides). PM: parabolic mirror. (Adapted from [5]; used with permission from OSA Publishing.) September 2019 37 Challenges for Biomedical Near-Field Imaging Near-field imaging can address the problem of lim- ited spatial resolution in a terahertz imaging system. A spatial resolution of 1 10 m n - seems a reasonable target for a pathologist. The tissue penetration and rapidity of the imaging process are other challenging factors needing to be addressed. The ability to pen- etrate tissue influences the guidelines for how big the clear margins of the normal tissue surrounding a can- cer tumor should be. Margins smaller than 1 mm may be adequate in some cases to protect against the recur- rence of ductal carcinoma in situ for women treated with lumpectomy [23]. Currently, the intraoperative response time for the simple surgical question of whether a lump is benign or malignant is 35–55 min, depending on whether the first frozen section was negative or positive. To speed up the imaging process, one needs to shorten the mea- surement times accordingly. Other challenges related to the tissue environment (i.e., pixel scanning and soft and uneven surface properties of the tissue) are discussed in the following. Soft-Tissue Imaging Freshly excised tissue exhibits a nonflat topography. The presence of blood and other fluids in the imaged regions may further influence the dielectric permittiv- ity measurement using a planar array of terahertz near- field sensors. Therefore, the concept is to use a flat, 2D array of sensors that can be pressed onto soft tissue. Pressing planarizes the tissue in the imaged region, and excess fluids are pushed aside, such that the sensors are brought in close proximity for margin identification. This requires a flip-chip sensor packaging concept where all interconnects are accessible on the chip’s back side, with through-silicon vias (TSVs) and back-side pat- terning of bond pad metallization. The Scanning Problem Future near-field imagers should exhibit 2D arrays with a sufficiently large number of pixels to image regions of interest in real time without the need for scanning or with a minimal scanning effort. Regular lateral scan- ning is simply impossible if the sensor array is in direct contact with soft tissue. The tissue will stick to the sur- face and cause stuttering during image acquisition. However, large-size tissues can be stepped incremen- tally, provided a large-pixel-count sensor is available. Breaking the Diffraction Limit The diffraction limit fundamentally restricts the spa- tial resolution of terahertz far-field imaging systems to the macroscopic scale, whereas numerous applications in biomedicine [24] or material characterization [25] require micrometer-range resolution. Breaking through this limit typically requires subwavelength-size aper- tures, scattering probe tips, or electro-optical probes to form an image along the scanned surface [2], [26]– [30]; this is known as near-field scanning optical micros- copy ( NSOM ). NSOM collects information about the 10 20 30 40 50 60 70 x -axis y -axis 10 20 30 40 50 60 70 10 20 30 40 50 60 70 x -axis y -axis (b) (c) (a) 10 20 30 40 50 60 70 Figure 4. An illustration of PCA over the entire FFT amplitude image. (a) A visible view of the sample. (b) The first principal component image. (c) The second principal component image in the 300–600-GHz band. (Adapted from [5]; used with permission from OSA Publishing.) Early work on biomedical terahertz near-field imaging primarily tried to understand which operation frequency is a good compromise among sensitivity, specificity, resolution, and system implementations. 38 September 2019 near-field interaction of a scanned object, with either an illuminated subwavelength aperture (A-NSOM) [30] or a scattering probe tip (S-NSOM) [25], as illustrated in Figure 5(a) and (b), respectively. Although very high spatial resolution, down to 20–40 nm, can be achieved [25], [27], [31], typical NSOM systems exhibit a low integration level and require long measurement times to reach a reasonable system sen- sitivity. The illumination or detection paths are placed remotely, resulting in extremely weak detection signals shadowed by strong far-field background clutter, mak- ing such imagers incapable of real-time visualization. Finally, all such bulky scanning techniques fail for soft and uneven tissue—considering the harsh environment during an oncology surgery. Integration of Terahertz Near-Field Sensors Several attempts have been made to increase the inte- gration of terahertz superresolution imagers with the goal of improved sensitivity and real-time imaging. The work in [30], [32], and [33] reported on a single-chip cointegration of subwavelength apertures with photo- conductive or cryogenically cooled electron-gas tera- hertz detectors; they demonstrated a 3 3 39 m - n - spatial resolution. Although the sensor was illuminated with high-power lasers, the signal dynamic range (DR) was still very low. An attempt to reduce image acquisition times was reported in [34], with a 14 m - n resolution around 0.7 THz, by combining a high-intensity source simultaneously illuminating a 370 740 m m - - # n n sample with an electro-optic lithium niobate crystal in sample proximity. Another effort to increase the number of subwave- length-size imaging pixels was presented in [35]. Here, a 4 4 # array of 15- n m-diameter apertures on a silicon wafer was developed to achieve a 30- n m spatial reso- lution at 2.54 THz. The array was illuminated from a high-power gas laser and a single hot electron bolom- eter detector at 4.2 K; a large collecting aperture was applied to capture the signals from all of the pixels. To differentiate among signals, frequency-division multi- plexing and the free-carrier absorption in silicon were used to modulate each aperture in the array. Silicon-Integrated Terahertz Near-Field Sensors Evanescent fields of resonant structures have been used for permittivity characterization of different materials. One approach is based on oscillator reac- tance sensors for permittivity characterization of bio- logical substances. In [36], a high-rate flow cytometer using oscillator reactance-based dielectric permittivity sensing was reported for frequencies between 6.5 and 30 GHz in 65-nm CMOS technology. In [37], a 12 16 # - pixel array of such sensors was reported at 120 GHz in 65-nm CMOS technology. However, the sensing ele- ments here are primarily composed of on-chip spiral inductors offering a spatial resolution of around 30 n m at the pixel level (an element size of m 90 160 m # n n in the array). Another popular approach is to characterize the trans- mitted or reflected wave at a loaded dielectric sensor. In [38], a permittivity sensor based on a 125-GHz reflec- tometer with an on-chip oscillator and interdigitated capacitor-based sensing element was reported. In [39], a 240-GHz permittivity sensor with a 30-GHz external local oscillator drive was shown in 130-nm SiGe hetero- junction bipolar transistor (HBT) technology. Here, an in- phase and quadrature-based coherent approach was used for complex permittivity sensing. The sensing element offers limited spatial resolution because it detects the change in phase-velocity characteristics of a rather long transmission line (TL), which is dielectrically loaded in its evanescent-wave field. The localization of this evanescent field to a narrow or confined area is needed to sense permittivity simultane- ously with a good spatial reso- lution. Also, because of a large circuit footprint, such a design is not suited for a dense array integration. Recent reviews on dielectric permittivity-sensing techniques for biomedical applications in silicon inte- grated circuits can be found in [40] and [41]. Terahertz Source Terahertz Source Aperture Vibration Sample Sample Confined E-Field Terahertz Detector Terahertz Detector B E Sample (a) (b) (c) Figure 5. Illustrations of (a) the A-NSOM technique, (b) the S-NSOM approach, and (c) SRR-based near-field sensing. B: B field; E: E field. (From [44].) September 2019 39 Recently reported silicon-integrated near-field sen- sors have achieved major improvements with respect to sensitivity and integration levels for terahertz super- resolution imaging [42]–[44]. The novel near-field sens- ing technique is based on the capacitive interaction of highly confined E-fields of the gap of a split-ring resonator (SRR) with the target material, as illustrated in Figure 5(c). Using this method, single-pixel sen- sors with a spatial resolution of 10 n m and an SNR of around 40 dB were achieved in [42] and [43]. However, such single-pixel sensors require raster scan- ning for image acquisition, which hampers their use in practical applications. An additional major imple- mentation challenge is building a highly confined sub- wavelength near-field probe in a planar technology, which is difficult in view of the thin, electrically small dielectric stack. Another issue arises from operation at terahertz frequencies, which restricts potential circuit architectures to very low complexity for sensor illu- mination and detection. Unlike NSOM, such sensors monolithically integrate the source, the evanescent- field sensor, and the detector on the same silicon die. This work has led to the first 128-pixel system on chip (SoC) for terahertz near-field imaging [44]. Terahertz Integrated Circuits for Near-Field Imaging Future applications will be massively impacted by the availability of advanced terahertz systems that can be adopted commercially and so profit from economies of scale. Thus, silicon integration is regarded as a key enabler for terahertz technology, providing compact and low-cost systems with large-scale mixed-signal integration. The European Union project DOTSEVEN [19], for instance, researched a 700-GHz f max SiGe technology, and recently CMOS technologies have achieved f T and f max of 485 GHz [45] and 350 GHz [46], respectively. Earlier research on silicon-based terahertz sys- tems was mainly focused on components for far-field imaging [7], including high-power radiation sources [47]–[49], heterodyne receivers [50], and direct power detectors [51], [52]. Integrated circuits can cointegrate all functions, such as illumination, evanescent-field sensing, and detection, in a tiny, room-temperature solid-state device. Additionally, current semiconduc- tor technologies can scale up to large-count imaging arrays able to reveal the microscopic nature of dynamic processes at terahertz frequencies in real time. Conceptual Design Terahertz sensing concepts should be simple enough to support efficient sensor illumination and detection. The idea applied in [42], [43], and [53] is to observe a shift in the resonance frequency of an SRR indirectly upon loading it with a sample. Instead of applying time-/ frequency-domain techniques to track the resonance frequency shift from f 0 to , f 0 l the SRR is simply illu- minated by a free-running on-chip oscillator at an off- set frequency f osc above or below f 0 The transmitted Transmitted Power 1 f osc f osc f 0 Source Evanescent-Field Sensor Detector Temporal Response Lucent Lucent Opaque Opaque x 2 Frequency f 0 ′ Figure 6. An illustration of the conceptual sensor operation principle. See [42] and [43] for implementation details. (From [42].) Magnetic Coupling 50 m Sensing Area Buried Split Gap W L S Top Metal C sense C sg C sg L 2 TL2 TL 2 TL 2 TL2 TL 1 TL 1 P2 L 2 L 1 k k P1 Metal 5–6 (a) (b) (c) Figure 7. An on-chip 3D SRR integrated into a multilayer BEOL stack [44]. (a) A 3D view. (b) A top view. (c) The resonator equivalent circuit. The resonator dimensions are , 4.5 W m n = , 3 S m n = and 15 L m n = TL: transmission line; p: port; C: capacitance; L: inductance; k: coupling factor. (From [44].) 40 September 2019 power through the SRR is detected with a broadband SiGe HBT power detector, as shown in Figure 6. In other words, the SRR acts like a tunable transmis- sion gate located in plane between the source and the detector. The frequency shift is translated into tem- poral transmission changes from opaque to lucent at the detector output. Note that, while more complex architectures such as an integrated heterodyne detec- tor may provide better sensitivity, they require too much area to facilitate multipixel integration and low dc power consumption. The detection of a sample-induced frequency shift across a resonant structure is an effective and com- monly applied technique exploited in various applica- tion scenarios in the low-frequency range and for terahertz frequencies as well. In particular, the interac- tion of near fields around SRRs with objects in direct contact or close proximity has been demonstrated for dielectric-permittivity and thickness sensors in the terahertz frequency band [54], [55]. Near-Field Sensor Implementation In a superresolution sensor, only a micrometer-range, spot-like sensing area should be exposed to the SRR surface, while the electromagnetic fields of the source and detection components need to be completely isolated from the top surface. The work presented in [42] solved this issue by exploiting the 15.3- n m- thick multilayer back end of line (BEOL) of a 0.13- n m SiGe BiCMOS technology for implementation of a 3D cross-bridged double SRR topology at 550 GHz. Fig- ure 7 shows the 3D topography and an equivalent cir- cuit for the SRR. Additionally, the sensor was excited by a buried TL, and the oscillator was implemented with striplines to avoid parasitic interactions with the imaging object. Most classical terahertz resonators on thin, low- permittivity dielectric substrates [55], [56] are illu- minated externally at normal incidence. An in-plane illumination challenges the resonator sensitivity and its lateral resolution because the electromagnetic fields are highly confined into a substrate and must be rearranged over a very short distance to produce a maximum evanescent-field strength above the tiny top-sensing surface [42]. The terahertz wave is magnetically coupled ( ) L , 1 2 with two shunt resonant structures compr i si ng microstrip lines closing the loops ( ) TL 2 and a buried split gap ( ) C sg used for sensor miniaturization. The sensing area is realized by two closely spaced lines in the top metal layer (C ) . sense The resonator exposes a highly confined electric dipole-type field to the top surface of the chip. The SRR operates in reflection mode, where the dominant object interaction is the bending of the E-field from the sensing volume and is related to the real part of the permittivity [42]– [44]. Therefore, both metallic and dielectric objects with a varying refractive index can be imaged and mapped to a monotonic permittivity-dependent sen- sor response, with the upper bound provided by met- als. The SRR shows a lateral resolution of 10–12 n m in the x and y directions and a sensitivity reduction of around 50% at a 1- n m distance. We refer readers to [42] for a discussion of the details and tradeoffs of the electromagnetic design of the SRR. Further improvements toward the resolution preferred by pathologists (1–10 n m) require other BEOLs with narrower linewidth and spacing constraints on the top-level metallization. Multipixel Near-Field Imager The primary challenge associated with multipixel near-field sensor integration is achieving dense imag- ing object coverage without dead zones between the sensing elements while simultaneously preserv- ing sensor sensitivity concerning pixel cross-coupling effects. In this regard, an operation frequency of arou nd 500 GHz offers a favorable comprom ise between available source power, detector noise equiv- alent power, and SRR footprint for the BiCMOS pro- cess technology. This frequency range is also a tradeoff between the permittivity-based image contrast of tis- sue and sensor resolution. However, a circuit archi- tecture where each SRR has a dedicated illumination source [42], [43] cannot be effectively applied to mul- tipixel arrays because of the size mismatch between the SRR and the oscillator. In general, on-chip tera- hertz oscillators require wavelength-scale TL systems in the top BEOL layers. With the SRR being of only subwavelength size, the oscillator dominates the total sensor footprint [42]. Therefore, multiple SRRs must SRR 50 m Power Splitter Oscillator 6.1 mm 1.75 mm 2.5-mm Distance to Bond Pads Refgen BP + Lock-In ADC ASIC Figure 8. An SoC micrograph (6.10 1.75 mm ) 2 # The total sensing area is 3.2 mm long. The blacked-out region contains circuits that are not related to the near-field array. ASIC: application-specified integrated circuit; ADC: analog-to- digital converter; Refgen: reference generator; BP: bandpass. (From [44].) September 2019 41 share a common oscillator for a dense SRR integration. Figure 8 shows the chip micrograph of a 550-GHz, 128- pixel near-field sensing SoC with cointegrated mixed-sig- nal baseband signal process- ing in a 0.13- n m SiGe BiCMOS tec h nolog y [53]. T he c h ip consists of 128 sensing pixels with individual cross-bridged double 3D SRRs arranged in a 3.2-mm-long 2 64 # 1D array. Besides sensor illumination, near-field sensing, and detec- tion, the chip is cointegrated with a readout integrated circuit for real-time image acquisition. The sensors are grouped into subarrays of four SRR–detector units and connected in paral- lel to the corresponding source element by means of a four-way equal-power divider. The SRRs in each group are arranged in a row with a 50- n m pitch, which is selected as a compromise between the fill factor and pixel cross coupling through the near field. The terahertz wave generated by a single oscilla- tor is fanned out to four SRRs, spaced at a 50- n m pitch. With further vertical mirroring and an extra 25- n m displaced pitch, the 1D stripe reaches an effec- tive 25- n m pitch and a fill fac- tor of 48%, corresponding to a density of around 100 pixels per inch (ppi). Array Readout Architectures Like visual cameras, terahertz multipixel sensor arrays can be coupled with an electronic global shut ter or a rolli ng shutter. With a global shutter, all pixels are operated simul- taneously and then read out sequentially. A rolling shutter instead captures each frame through rapid electronic scan- ning across the array. The lat- ter method may be slower, but Supply Voltages PC FPGA Real-Time Visualization/ Data Logging UART (USB) SPI Clocks row[0] chop OSC ref row[1] chop OSC ref row[31] 5 row[31:0] ASIC LIgain[6:0] detbias[5:0] BPgain[5:0] adcout[5:0] refctrl[43:0] VDD VDD det VCC osc OSCref GND adc clk col[3:0] SPI[4:0] chop OSC ref 44 6 6-b FADC V refp V refm Reference Generator TP2 TP1 Chip Boundary Active Load BP ref LP clk LI ref chop Lock-In Amplifier BP Amplifier 7 6 6 v det vref DETbias Power Detector Power Detector Power Detector SRR SRR SRR 1:4 Power Splitter 1:4 Power Splitter 1:4 Power Splitter row[0] row[1] row[31] col[3:0] col[3:0] col[3:0] TPO TPO TPO Figure 9. A block diagram of a 128-pixel near-field SoC [53], including external components. FADC: flash analog-to-digital converter; GND: ground; LIref: lock-in reference; OSC: oscillator; TP: test point; UART: Universal Asynchronous Receiver/ Transmitter; VDD: digital supply voltage; VCC: analog supply voltage. (From [53].) 42 September 2019 it benefits from lower power consumption, as only one sensor of the array is activated at a time. The shutter speed corresponds to the exposure time of the sensor. A required readout rate of 28 frames per second (fps) of 128 pixels, for instance, requires a maximum integra- tion time of about 300 n s per pixel. The terahertz direct detectors of a sensor pixel are read out either after amplification at dc or with the help of a lock-in ampli- fier at a small ac reference frequency offset (chopping frequency). The 1,024-pixel terahertz camera in [51], for instance, is read out at dc, whereas the 128-pixel array in [53] uses an on-chip lock-in amplifier readout instead. Lock-in amplifiers use phase-sensitive detection to single out the component of the signal at the chopping frequency. Noise frequencies other than the reference frequency (e.g., low-frequency / f 1 noise components) are rejected and do not affect the measurement. Figure 9 shows the block diagram of the 128 near- field pixels with an on-chip lock-in amplifier readout scheme. The array consists of two 64-pixel rows of SRRs and power detectors. Each row is divided into 16 subar- rays of four pixels, which are driven concurrently from an individual chopped triple-push oscillator (TPO). All pixels share a global readout chain comprising an active PMOS load, a bandpass (BP) amplifier, a lock-in amplifier, and a 6-b flash analog-to-digital converter. The array is operated sequentially with a rolling shut- ter, where only one TPO and one detector are turned on at a time. A bandgap-based reference generator is used to provide process-voltage temperature-stable reference voltages on the chip. Clocking, pixel address- ing, and the programmable settings (for amplifier gains and reference voltages) are all controlled by a cointe- grated, application-specified integrated circuit, which additionally includes a Serial Peripheral Interface (SPI) slave communicating with an external controller, such as a field-programmable gate array (FPGA). The chip features both an analog readout mode and a lock-in amplifier-based digital readout mode. Packaging and Sensor Module Design Sensors for biomedical applications, such as real-time tumor assessment, must be able to handle samples that are both wet and soft. To facilitate this, appropri- ate encapsulation is needed.