Recent Advances in Ultrasound and Photoacoustic Analysis for Thyroid Cancer Diagnosis

Thyroid cancer is one of the most commonly diagnosed cancers worldwide, with a continuously increasing incidence rate in recent decades. Although ultrasonography, which is the current screening method in clinical workflows, has successfully triaged cancerous nodules for biopsy, overdiagnosis has also grown due to the relatively low specificity of the method. Studies are conducted to overcome this overdiagnosis issue by complementing ultrasonography with additional image‐based analysis techniques. This review presents an overview of the current advances in clinical trials using advanced ultrasound (US) and photoacoustic (PA) imaging techniques for thyroid nodules in humans. A summary of initial trials by Doppler US and US elastography to improve the classification accuracy of thyroid nodules is presented. Furthermore, recent PA techniques with multispectral analyses utilizing clinically available machines are explored. By amending the existing ultrasonography, the advanced US and PA techniques can enhance the triaging accuracy by analyzing both structural and functional information of thyroid nodules in vivo.


Introduction
Thyroid cancer is one of the most common cancers in the world. [1][2][3] In 2018, the age-standardized rates per 100 000 personyears were 10.2 and 3.1 for women and men, respectively. [4] Compared to other commonly diagnosed cancers, such as breast, prostate, lung, and colorectal cancers, the death rate of thyroid cancer is relatively low; however, the incidence rate has been continuously increasing over several decades. [5][6][7] According to the American Cancer Society, ≈43 800 newly diagnosed cases of (US)-based signal processing. [15][16][17][18] Another technique for providing molecular information about nodules is photoacoustic imaging (PAI), which integrates USI and optical imaging. [19][20][21] In this review article, an overview of the principles and characteristics of USI and PAI is provided. Recent advances in these techniques for analyzing thyroid cancer are summarized herein. The classification results of the image-based analysis techniques show great potential for clinical application. By complementing the existing USI-based risk stratification with structural and functional information, the triaging accuracy of nodules can be enhanced.

Principles of Imaging
USI is one of the most widely used medical imaging techniques in clinics. [22][23][24] It is advantageous for routine examinations because it is non-ionizing, cost-effective, real-time imaging, and easy to use compared to other conventional imaging techniques such as X-rays, CT, MRI, and positron emission tomography (PET). US images can visualize the anatomic structures in a cross-sectional view, which are called B-mode images. The principle of USI is the transmission and reception of acoustic waves using various US transducer geometries. [25] The transmitted acoustic waves physically interact with the internal structures in the body, and the reflected waves are then propagated to the original position (Figure 1a). Because the images can be obtained in real-time (typically 30-60 fps), instantaneous guidance for medical procedures is possible. [26] However, conventional USI is incapable of obtaining the functional information required to differentiate lesions from healthy tissues.
Color Doppler imaging was used to identify the blood flow to add functional information to US data ( Figure 1b). Color Doppler imaging is based on the Doppler effect, in which the measured frequency is shifted by the relative velocity between the object and the detector. [27] In the USI, the measured frequency (f r ) is given by the following equation where v i is the relative velocity between the receiver (US transducer) and the source of the acoustic reflection (internal objects), v s is the speed of sound (1540 m s −1 in soft tissue), and f t is the frequency of the transmitted US waves. Then, the velocity of an internal object, which is typically blood flow, can be derived using the following equation where v is the speed of the blood flow, f d = f t − f r is the Doppler frequency shift, and is the angle of the moving direction with respect to the axial direction of the US transducer.
Recently, elastography-based imaging techniques have received attention for the assessment of cancerous nodules in clinics. [28] This technique visualizes tissue stiffness by calculating the modulus of elasticity using the measured deformation of the tissues in accordance with the applied mechanical forces (Figure 1c). Using Hooke's law, the elastic modulus (E) can be described by the following equation where = F/A is the stress, which is given by the force (F) per unit area (A), and = ΔL/L is the strain, which is the expansion (ΔL) per unit length (L). Several US elastography techniques have been developed using various excitation methods such as strain imaging and shear-wave imaging. [29] Measuring tissue elasticity is advantageous for providing qualitative information for diagnostic applications including the assessment of hepatic fibrosis, [30] renal lesions, [31] breast masses, [32] prostate cancers, [33] thyroid cancers, [34] and tendons. [35] In addition to the physical and mechanical movements in biological tissues, molecular functional information can also be obtained using the PAI technique by combining the principles of US and optical imaging (Figure 1d). The principles of PAI are based on the photoacoustic (PA) effect, which is the energy transduction from light to sound. [36][37][38] In the PAI procedure, laser pulses (typical pulse widths of a few nanoseconds) are used to illuminate the target objects. Light energy is absorbed by the objects and then released as heat, which is quickly dissipated. The rapid heat change generates a rapid volumetric change in the object, producing acoustic waves (i.e., PA waves). PA waves can be detected and reconstructed using US machines and algorithms. [39][40][41] In addition to visualizing intrinsic chromophores (oxyhemoglobin, deoxyhemoglobin, melanin, and lipids), [42][43][44] contrast-enhanced PA images can also be acquired by injecting optically absorbing agents. [45][46][47][48][49][50] Using multiple www.advancedsciencenews.com www.advphysicsres.com wavelengths, PAI can resolve molecular components in biological tissues using spectral unmixing techniques. [51][52][53][54] PAI can provide functional and pathological information on biological tissues from intrinsic biological chromophores. [55,56] 3. Ultrasound Analysis of Thyroid Nodules

Ultrasound-Based Risk Stratification
In the current clinical workflow, USI-based risk stratification is the gold standard method for triaging thyroid nodules. [57] The typical methods for risk stratification are the American Thyroid Association (ATA) Guideline, [58] the British Thyroid Association (BTA) Guideline, [59] and the Thyroid Imaging Reporting and Data System (TI-RADS). [60] Using the US images, which show the anatomical structure of nodules in vivo, stratification scores are assigned by clinicians according to the guidelines related to the structural features of thyroid nodules including size, shape (taller-than-wide), presence of calcifications, irregular margins, hypoechogenicity, and composition (cystic, solid, or mixed). [61,62] For the definitive decision, FNAB is performed on the nodules that satisfy the suspicious criteria. The representative classification criteria and diagnostic performances of the USI-based risk stratification methods are summarized in Table 1. [63][64][65] The USIbased risk stratification methods successfully triage malignant nodules with high sensitivity. However, they suffer from low specificity, leading to overdiagnosis, which can lead to unnecessary biopsies or surgeries.

Doppler Ultrasound Imaging
Doppler US images were used to assess the surrounding vascularity to add functional information for diagnosing thyroid nodules ( Table 2). Appetecchia et al. attempted a US-Doppler combined risk stratification method for distinguishing thyroid nodules. [66] They acquired Doppler US images and conventional US B-mode images from 203 patients (27 malignant and 176 benign). Using the Doppler US images, the patients were classified based on their vascularity around the nodule. The authors attempted to assess classification accuracy by combining vascularization factors with two types of US features, namely, the absence of a halo sign and microcalcification in the nodules. The combination of the two US features and intra-nodular vascularity showed a more balanced classification accuracy (72.2% sensitivity and 77.2% specificity), compared to the combination of one of the two US features (with the absence of a halo sign, 83.3% sensitivity, and 43.7% specificity; with microcalcification, 80.6% sensitivity, and 75.4% specificity). However, the enhancement of the classification accuracy was marginal. Moon et al. investigated the effectiveness of Doppler US responses in predicting malignant nodules. [67] They acquired data from 1083 nodules (269 malignant and 814 benign) and compared the diagnostic performance of US vascularity combined with that of conventional US features. The results showed that the addition of vascularity information (90.3% sensitivity and 56.3% specificity) was not as useful as the use of gray-scale US features alone (84.4% sensitivity and 85.9% specificity). Rosario et al. reported a similar result using 1502 thyroid nodules (160 malignant and 1342 benign). [68] In their study, the combination of vascularity and US features had no extra utility in predicting malignant nodules (89.4% sensitivity and 66.4% specificity) as compared to the US features alone (88.7% sensitivity and 68.2% specificity). In addition, Maddaloni et al. reported that Doppler US did not improve the diagnostic accuracy of the conventional US-based risk stratification method. [69] Recently, Zhu et al. demonstrated a machine learning-based technique for enhancing diagnostic performance using US Doppler data acquired from 712 thyroid nodules (356 malignant  and 356 benign). [70] They designed an artificial neural network model (TDUS-Net) by extracting features from US Doppler images (whole ratio, intranodular ratio, peripheral ratio, and the number of vessels) and US features defined by TI-RADS. For comparison, an additional artificial neural network model (TUS-Net) based on the US features alone was also designed. The results showed improved classification accuracy for TDUS-Net (79.2% sensitivity and 89.9% specificity) compared to TUS-Net (77.8% sensitivity and 87.8% specificity 87.8%). Interestingly, these results were better than the classification accuracy performed by the radiologists (68.2% sensitivity and 79.5% specificity). The results showed that the machine learning technique could improve diagnostic performance by using vascularity information obtained from Doppler US images.

Ultrasound Elastography
US elastography has also been used to improve the diagnosis of thyroid cancer (  for thyroid nodules according to their visibility, brightness, regularity, and margin distinction. In addition, the thyroid-to-tumor strain ratio (strain index) was calculated using offline processing.
In a previous study, Park et al. used US shear waves as an external pressure for elastography to measure quantitative tissue stiffness. [72] Combinations of US features and elasticity indices were evaluated for 476 thyroid nodules (379 malignant and 93 benign). The combined criterion showed better classification performance (94-95% sensitivity and 34-56% specificity) compared to the conventional US features (92.9% sensitivity and 60.8% specificity) and elasticity indices (34-54% sensitivity and 60-88% specificity). However, the physiological motion of patients during data acquisition degrades the reproducibility of classification results.
Dighe et al. proposed US elastography using the natural pulsation of the carotid artery as a pressing source to overcome the displacement of patients. [73] They recruited 53 patients (10 papillary carcinomas, 25 follicular lesions, 12 multinodular goiters, 4 thyroiditis, and 2 Hürthle lesions) to evaluate their approach. The results showed a promising ability to differentiate papillary carcinoma (87.8% sensitivity and 77.5% specificity) from other types of thyroid cancer and benign nodules. Using the same technique, Lim et al. evaluated the inter-observer agreement and intra-observer reproducibility of 56 thyroid nodules (53 malignant and 3 benign). [74] Recently, Zhao et al. demonstrated a machine-learning technique for analyzing diagnostic performance using a dataset of USI and shear wave elastography. [75] They acquired data from 849 thyroid nodules (307 malignant and 542 benign nodules). The nodules were divided into a training set (192 malignant and 328 benign), a validation set (82 malignant and 141 benign), and a testing set (33 malignant and 73 benign) to train and validate the decision models. Two models were trained to evaluate the usefulness of shear wave elastography. The first model was trained using five US features, which were internal composition, echogenicity of the solid portion, shape, margin, and echogenic foci. The second model was trained using five features of shear wave elastography (i.e., mean value, minimum value, maximum value, standard deviation, and ratio of mean values between nodules and parenchyma), in addition to the five US features. The second model showed better classification performance (93.9% sensitivity and 93.2% specificity) than the first model (90.9% sensitivity and 78.1% specificity) and conventional TI-RADS risk stratification (90.9% sensitivity and 45.2% specificity).

Ex Vivo Photoacoustic Analysis
At the initial stage, ex vivo tissues were used to test the feasibility of PA analysis for investigating thyroid nodules ( Table 4). Kang et al. investigated ex vivo thyroid microcalcification using PAI and hypothesized that thyroid microcalcification is correlated with thyroid malignancy. [76] The system was configured with a linear transducer of 5-14 MHz and a laser of 10 Hz rate. The wavelength and energy of the laser were set at 700 nm and 10 mJ cm −2 , respectively. The study included 36 resected thyroid nodule samples obtained immediately after surgery from 18 patients. PAI results were compared with radiographs, USI, visual examination, and histological analysis. However, the results did not yield any relevant outcomes from microcalcifications in PAI, whereas a significant difference in intensity (P = 0.007) was observed in the US images and radiographs. The authors emphasized the limitations of the single-wavelength analysis and the remaining blood from ex vivo tissues as reasons for the inability to distinguish microcalcifications with PAI.
Sinha et al. investigated multispectral PAI for ex vivo thyroid analysis to overcome the limited information of singlewavelength PAI. [77] They performed a frequency analysis of multispectral PA signals to differentiate malignant nodules from normal tissues. They used a 32-element linear transducer with a center frequency of 5 MHz and a tunable laser at a rate of 10 Hz. The target tissues were fixed in a custom-designed PA imaging probe (Figure 2a). Wavelengths of 760, 800, 850, 930, and 970 nm were used to obtain multispectral PA signals, which correspond to the peak absorption coefficients of deoxyhemoglobin, hemoglobin, oxyhemoglobin, water, and lipids, respectively. The laser fluence was maintained at 5 mJ cm −2 for all the wavelengths. Multispectral PA images were obtained from six excised thyroid tissues (Figure 2b). The malignant region (blue circles in Figure 2b) showed distinguishable signals in the PA images. Frequency analysis was performed on the normal and cancerous regions from the acquired multispectral PA images. The calibrated power spectrum of the PA signals at each wavelength was fitted to a linear model, and three parameters (slope, mid-band fit, and intercept) were quantified. The results showed that the mid-band fit Dorga et al. demonstrated their algorithm to calculate the distribution of individual chromophores in multispectral PA images. [78] They acquired multispectral PA images from 88 excised thyroid sections using wavelengths of 760, 850, 930, and 970 nm, which correspond to deoxyhemoglobin, oxyhemoglobin, lipid, and water, respectively. The average intensity of each chromophore was calculated. The mean intensity of spectrally resolved deoxyhemoglobin in malignant thyroid tissue was statistically different from that in nonmalignant thyroid tissue (P = 0.011, sensitivity 69.2%, specificity 96.9%).

In Vivo Photoacoustic Analysis
To evaluate PA analysis techniques for clinical applications, in vivo studies were also performed ( Table 5). Dima et al. demonstrated the feasibility of in vivo human thyroid imaging by developing a hand-held imaging probe using a 64-element curved array transducer with a center frequency of 7.5 MHz. [79] The optical fiber bundle was mounted on the transducer and designed to adjust the angle of the fiber. A single wavelength of 800 nm with a repetition rate of 10 Hz was used. The fluence was maintained below 20 mJ cm −2 . Two healthy volunteers were recruited, and in vivo PA images were acquired. The results demonstrated the feasibility of PAI for visualizing blood vessels with a better resolution than US Doppler.
Yang et al. developed a dual-modal PA and US imaging system to visualize the anatomical structure and PA distribution in malignant nodules. [80] A 192-element linear transducer with a center frequency of 5.8 MHz was used with a 1064 nm laser at a rate of 10 Hz. The laser fluence was 5.5 mJ cm −2 at the skin surface. A custom holder was made, and two fiber optic bundles were mounted on either side of the transducer (Figure 3a). The overlaid PA and US images visualized the thyroid glands and surrounding tissues of three healthy volunteers and ten malignant patients (Figure 3b). For subsequent validation, US color Doppler  images were acquired and compared with the PAI results (inset in Figure 3b). PAI detected more vascular signals in both the nodule and its surrounding thyroid parenchyma, whereas US Doppler could not visualize the peripheral vasculatures. However, the system is limited to clinical studies due to the lack of multispectral and real-time PAI. Mercep et al. introduced a multi-segmented US transducer for multispectral PAI of the thyroid region. [81] The multi-segmented US array consisted of a 128-element linear segment and two 64element concave segments with a center frequency of 7.5 MHz (Figure 4a). The laser was operated at a repetition rate of 25 Hz with a maximum fluence of 10 mJ cm −2 . For multispectral imaging, seven optical wavelengths (700, 730, 760, 780, 800, 825, and 850 nm) were used. All 256 elements in the multisegmented array were used for PA signals, whereas only the linear segment was activated for the US signal. To test the feasibility of the system, multispectral PA and US images were acquired from a healthy volunteer (Figure 4b-d). The spectrally unmixed oxyhemoglobin and deoxyhemoglobin maps showed higher levels of oxyhemoglobin signals in the carotid artery, whereas deoxyhemoglobin signals were more pronounced in the jugular vein and subcutaneous microvasculature. Although this study was conducted on only one healthy volunteer, spectral unmixing was applied to observe the distribution of oxyhemoglobin and deoxyhemoglobin in the neck tissues including the thyroid gland.
Roll et al. evaluated malignant and benign thyroid nodules in vitro to determine the effectiveness of multispectral PA and US imaging. [82] They used a 256-element linear array transducer with a center frequency of 3 MHz (Figure 5a). Multispectral PA images were acquired using eight different wavelengths (700, 730, 760, 800, 850, 900, 920, and 950 nm) with a repetition rate of 25 Hz and fluence of <20 mJ cm −2 . The multispectral PA properties of 13 normal and 3 malignant nodules were analyzed. They delineated the region of interest (ROI) on the US image and transferred it to the corresponding co-registered oxygen saturation (sO 2 ) map ( Figure 5b). The statistical analysis showed a significant difference in sO 2 levels between malignant and benign nodules with a P-value of 0.0393 (Figure 5c). The lipid content of malignant nodules was lower than that of normal nodules, but the difference was not statistically significant (P = 0.1295). This study is significant in that it statistically distinguished between malignant and benign nodules in vivo.
For a more robust clinical study, Kim et al. proposed a novel metric to distinguish malignant thyroid nodules by combining machine-learning-powered multiparametric PA analyses with standard US risk stratification. [83] A dual-modal PAUS imaging system (Figure 6a), composed of a programmable clinical US machine and a 10 Hz-rate laser source with a fluence of 10 mJ cm −2 , was used to obtain data from 52 patients (23 malignant, 29 benign). A hand-held imaging probe composed of a 128-element linear array transducer with a center frequency of 8.5 MHz and a laser-delivering fiber bundle was designed for facile data acquisition in humans (Figure 6b). Multispectral PA images were acquired using five different wavelengths: 700, 756, 796, 866, and 900 nm (Figure 6c). From the nodule boundaries determined by clinicians, three parameters (sO 2 level in the nodule, spectral gra-  dient of the PA signal, and skewness of the sO 2 distribution in the nodule) were quantified from the multispectral PA responses. A decision boundary was determined using a machine learning technique using the three parameters, and its classification accuracy (78% sensitivity and 93% specificity) was demonstrated. To improve the accuracy even further, a new risk stratification (ATAP score) was proposed by combining the PA analysis results with the conventional ATA guideline score. This new metric could distinguish malignant from benign nodules with a sensitivity of 83% and a specificity of 93% (Figure 6d). This is the first clinical study with a sufficient number of patients for statistical analysis.

Conclusion
Successful screening of thyroid nodules should achieve high specificity to minimize the overdiagnosis of normal thyroid nodules while maintaining high sensitivity for malignant nodules. Non-invasive and non-ionizing PAI techniques that combine high optical contrast and acoustic resolution have advanced over the past few years to identify malignant thyroid nodules. This review summarizes current advances in US and PA analyses for enhancing diagnostic accuracy using functional information on thyroid nodules.
The major approaches for obtaining functional information from US pictures were vascularity from Doppler US and stiffness from US elastography. Internodular vascularity has been combined with conventional US features in an attempt to improve classification accuracy. However, recent studies have shown that the correlation between vascularity and malignancy is marginal in thyroid nodules. Nodule stiffness was quantified to provide additional functional evidence of malignancy. The results showed improved classification accuracy; however, reproducibility and stability are key issues for successful clinical application. Recently, machine-learning techniques have been used to enhance diagnostic accuracy by training a decision model using vascularity or stiffness with conventional US features. These results show great potential for better classification results.
PA thyroid nodule studies were categorized from ex vivo to in vivo, from single-wavelength PAI to multispectral PAI, and from single PAI to multimodal PAUSI. This is the result of the unremitting efforts of researchers around the world to propose meaningful results. This review describes two major advances in PAI systems for distinguishing between malignant and normal thyroid nodules: 1) Multispectral PAI allows the extraction of multiple endogenous chromophores from the thyroid tissue. This solves problems such as the background signal and the limited parameters of a single-wavelength PAI. 2) Multimodal Figure 6. Multiparametric PA analyses of the malignancy of thyroid nodules in vivo. a) Photograph of the clinical PA and US imaging system, b) photograph and schematic of the imaging probe for dual-modal PA and US imaging, c) representative overlaid PA and US images of the thyroid region from benign and malignant patients in vivo, and d) classification result using a new risk stratification (ATAP score). PA, photoacoustic; US, ultrasound; TR, transducer; FB, fiber bundle; sO 2 , oxygen saturation level; TH, thyroid lobe; CA, carotid artery; ND, nodule; PTC, papillary thyroid cancer; B, benign; Se, sensitivity; Sp, specificity. Reproduced with permission. [83] Copyright 2021, American Association for Cancer Research.
PAUSI helped physicians familiar with USI obtain accurate PA images by minimizing the sense of heterogeneity.
Despite these remarkable advances, the clinical application of PAI to thyroid nodule screening still faces significant challenges. First, although PAI achieved a deep imaging depth, even though it is a light-based technology, the degradation of the signal-tonoise ratio (SNR) owing to the penetration depth of light in tissue remains a challenge. In particular, because the depth of the thyroid nodule may differ by several centimeters or more for each patient, the SNR of the PA signal according to the depth must be constant for accurate analysis. This may be overcome by conducting more research on the light transmission mechanism, US probes of various geometries, and optimal combinations of US probes and optical fibers. [84][85][86] The second is the trade-off between the practicality and accuracy of portable PAI systems. Because PAI can be easily combined with conventional USI sys-tems, most systems use small handheld probes. [87][88][89][90] However, these portable probes may not be able to photoacoustically represent the actual anatomy of thyroid tissue because of their limited angle of view, which can possibly hinder the accurate interpretation of PA signals. Deep learning techniques [91][92][93][94][95][96] or new PA image reconstruction techniques [97][98][99][100][101] may be the key to overcoming these problems. Finally, standardization of PAI systems is required. PAI systems have a very wide range of components (e.g., lasers, transducers, image reconstruction methods, and display methods), even for the same purpose or application. This can lead to different interpretations of the same lesions. This urges global researchers to put more effort into developing standardized systems from the perspective of end-users (i.e., physicians). Finally, overcoming this obvious barrier will allow for the clinical translation of thyroid nodule examination using PAI.