452

IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 17, NO. 2, MARCH 2013

The Tactile Sensation Imaging System for Embedded Lesion Characterization Jong-Ha Lee and Chang-Hee Won

Abstract—Elasticity is an important indicator of tissue health, with increased stiffness pointing to an increased risk of cancer. We investigated a tissue inclusion characterization method for the application of early breast tumor identification. A tactile sensation imaging system (TSIS) is developed to capture images of the embedded lesions using total internal reflection principle. From tactile images, we developed a novel method to estimate that size, depth, and elasticity of the embedded lesion using 3-D finite-element-modelbased forward algorithm, and neural-network-based inversion algorithm are employed. The proposed characterization method was validated by the realistic tissue phantom with inclusions to emulate the tumors. The experimental results showed that, the proposed characterization method estimated the size, depth, and Young’s modulus of a tissue inclusion with 6.98%, 7.17%, and 5.07% relative errors, respectively. A pilot clinical study was also performed to characterize the lesion of human breast cancer patients using TSIS. Index Terms—Elasticity imaging, mechanical imaging, optical imaging, tactile imaging system, tactile sensor, tumor detection.

I. INTRODUCTION CCORDING to the American Cancer Society, more than 178 000 women and 2 000 men are found to be afflicted with breast cancer every year; international statistics report an estimated 1 152 161 new cases annually [1]. This form of the disease is the leading killer of females between 40 and 55 years of age, and is statistically the second leading cause of death overall in women [2]. Clearly, early detection and diagnosis is the key to surviving this fatal disease. In order to detect various forms of breast cancer, there are many methods used today such as computer tomography (CT), ultrasonic imaging (US), magnetic resonance imaging (MRI), and mammography (MG) [3]–[6]. However, each of these techniques has disadvantages: harmful radiation to the body (CT, MG), low specificity (MRI), complicated system (MRI), low image resolution (US), etc. The criteria for an efficient early breast cancer detection system include: accuracy, high sensitivity, having acceptable specificity, easy to use, acceptability in terms of levels of discomfort and time taken to perform the test, and cost effectiveness.

A

Manuscript received May 6, 2012; revised December 7, 2012; accepted January 25, 2013. Date of publication February 5, 2013; date of current version March 8, 2013. J.-H. Lee (corresponding author) is with the Department of Biomedical Engineering, School of Medicine, Keimyung University, Daegu 704-701, Korea (e-mail: [email protected]). C.-H. Won is with the Department of Electrical and Computer Engineering, Temple University, Philadelphia, PA 19122 USA (e-mail: [email protected]). Color versions of one or more of the figures in this paper are available online at http://ieeexplore.ieee.org. Digital Object Identifier 10.1109/JBHI.2013.2245142

During the past two decades, various methods have been devised for measuring or estimating the tissue stiffness. Elastography is one of the actively researched elasticity determination methods [10], [11]. If compression or vibration is applied to the tissue, the embedded tumor deforms less than the surrounding tissue due to its high stiffness characteristics. Under this observation, elastography records the distribution of tissue elasticity. Elastography is currently performed using ultrasonic, magnetic resonance, and atomic force microscopy [12]–[14]. While elastography is successfully applied to static organs such as liver, breast, and brain, the calculation of tissue inclusion stiffness is still challenging. Also the instrument is relatively expensive and requires a dedicated operator [15]–[17]. Recently, a new technological method, “elasticity imaging using tactile sensor” has been explored [18]–[21]. This type of technology calculates tissue elasticity by sensing mechanical stresses on the surface of tissues using tactile sensors [22], [23]. The medical device named “SureTouch Visual Mapping System” produced by Medical Tactile Inc., is one of the elasticity imaging system using tactile sensors [24]. The device consists of a probe with capacitive pressure sensor arrays and electronic units to transmit tactile data to the computer. Using a 32×32 tactile sensor array, the device obtains the stress distribution on the tissue surface. The device is capable of computing and visualizing the pressure pattern of the tissue. However, the resolution of pressure sensor-based method is not as good as optically-based method. Also the device requires other sensors to detect the applied force. We developed a tactile sensation imaging system (TSIS) based on the total internal reflection (TIR) principle. Light is injected into the sensing probe of the system to allow for TIR. When the embedded lesion is compressed from the surface by the system, the trapped light in the sensing probe is scattered. The scattered light is captured by a high resolution near-infrared camera. These captured images provide maximum pixel values, total pixel values, and indentation area. From these images, we developed a novel method to estimate that size, depth, and elasticity of the embedded lesion. The benefits of the proposed TSIS over the existing techniques are nonionizing, cost effective, and easy to access over the existing techniques. To estimate tissue inclusion parameters using tactile sensor, different approaches have been explored. In [25], finite-element modeling (FEM)-based forward algorithm and Gaussian fitting model-based inversion algorithm are devised. This study was extended in [26] to attempt to find a more complete set of tissue inclusions. They showed that the estimation results are more accurate in determining the size of a tissue inclusion than manual palpation. Nevertheless, the results are limited to tissue

2168-2194/$31.00 © 2013 IEEE

LEE AND WON: THE TACTILE SENSATION IMAGING SYSTEM FOR EMBEDDED LESION CHARACTERIZATION

inclusions at least 100 times stiffer than the surrounding tissues. In addition, other tissue inclusion parameters such as depth and Young’s modulus are not available. In [27], FEM-based forward algorithm and transformation matrix-based inversion algorithm were proposed to estimate size, depth, and Young’s modulus of the tissue inclusion. But the error in estimating tissue inclusion modulus is larger than our case. The main difference between [27] and our method is that we use 3-D FEM instead of 2-D FEM for the better forward algorithm accuracy. In addition, instead of transformation matrix-based inversion algorithm, we use an artificial neural network (ANN)-based inversion algorithm to increase the inversion algorithm accuracy. Finally, we use newly developed elasticity measurement system. A noninvasive method for estimating the elasticity of the tissue inclusion would offer great clinical utility. In addition to the stiffness, geometric parameters such as size and depth of the tissue inclusion are also important factors in assessing the tumor. The combined knowledge of stiffness and its geometry would aid in discerning the malignancy of the mass. In this paper, we present a novel method to estimate the stiffness and geometric information of a tissue inclusion. The estimation is performed based on the tactile data obtained at the tissue surface. To obtain the tactile data, we developed an optical TSIS. The optical method will have a better resolution and contrast compared to the ultrasound-based elastography. In order to capture the tactile data presented by a tissue inclusion, TSIS utilize the TIR principle. To estimate various tissue inclusion parameters— size, depth, Young’s modulus—using TSIS tactile data, the FEMbased forward algorithm and ANN-based inversion algorithm are developed. The FEM is used to generate simulated tactile data at the tissue surface over different tissue inclusion parameters in the idealized tissue model. The ANN maps the tactile data to the tissue inclusion parameters. To simulate a realistic tissue phantom, we used the 3-D FEM. For the ANN-based inversion algorithm, the scaled conjugate gradient algorithm (SCGA) is utilized as a learning algorithm due to its outstanding performance compared to the conventional learning algorithm such as resilient back-propagation algorithm [28]. Finally, to verify the proposed method, the realistic tissue phantom experiments were performed. Small-scale breast cancer patient experiment results are also presented. In the following section, the design concept of TSIS is introduced. Then the tactile sensation imaging principle and its verification using numerical simulation are discussed. Next, the tissue inclusion parameter estimation method is discussed. Then the proposed estimation method is validated using the realistic tissue phantom with different inclusion parameters. Finally, the conclusions and discussions are presented.

II. SYSTEM DESIGN CONCEPTS In this section, we present the design concept of TSIS. Then the tactile sensation imaging principle based on the TIR phenomenon is discussed. We performed numerical simulations to verify the principle behind TSIS. Finally, we show that we were able to obtain tactile images of phantom tissue inclusions.

453

A. System Overview The TSIS incorporates an optical waveguide unit, a light source unit, a high-resolution camera unit, and a computer unit. The optical waveguide is the main sensing probe of the system. The waveguide is composed of polydimethylsiloxane (PDMS), which is a high-performance silicone elastomer [30]. In the current system design, the waveguide needs to be flexible and transparent and PDMS meets this requirement. We emulate the tissue structure of the human finger. The human finger tissue is composed of three layers with different elastic modulus [31]. They are the epidermis, dermis, and subcutaneous. The epidermis is the hardest layer, with the smallest elastic modulus, and it is approximately 1 mm thick. The dermis is a softer layer, and it is approximately 1–3 mm thick. The subcutaneous is the softest layer which fills the space between the dermis and bone. It is mainly composed of fat and functions as a cushion when the load is applied to the surface. Due to the difference in hardness of each layer, the inner layer deforms more than the outer most layer when the finger presses into the object. To emulate this structure, three PDMS layers with different Young’s modulus are stacked together. The PDMS #1 layer is the hardest layer, PDMS #2 layer is the layer with medium hardness, and PDMS #3 layer is the softest layer. The height of each layer is approximately 2 mm for PDMS #1 layer, 3 mm for PDMS #2 layer, and 5 mm for PDMS #3 layer, respectively. The high resolution camera is a complementary metal–oxide semiconductor (CMOS) camera with 8.4 μm × 9.8 μm individual pixel size (Allied Vision Technology, Germany). It has a pixel array of 768 H × 492 V with 8-bit depth, thus its maximum resolution is 0.4 megapixel. The camera is placed below a waveguide. A heat-resistant borosilicate glass plate is placed between the camera and waveguide. The internal light source is a white lightemitting-diode (LED) with a diameter of 1.5 mm. There are four LED light sources placed on four sides of the waveguide to provide illumination. The direction and incident angle of light is calibrated to be totally reflected within the waveguide. B. Total Internal Reflection The proposed TSIS operates on the principle of TIR. According to Snell’s law, if two mediums have different refractive indices, and the light is shone through those two mediums, then a fraction of light is transmitted and the rest is reflected [32]. If the incident angle is above the critical angle, then TIR occurs. In the current system design, since the waveguide is surrounded by the air, and has a lower refractive index than PDMS layers, the incident light directed into the waveguide is totally reflected in the waveguide. The waveguide is transparent and flexible. Consequently, if a waveguide is compressed by an external force toward a stiff inclusion, the contact area of the waveguide deforms and causes the light to scatter. The scattered light is then captured by the high-resolution camera and saved as an image. We process these tactile images to estimate the mechanical properties of the lesions. Thus, the basic principle of tactile sensation imaging lies in capturing of the light scattered due to the inclusion. Fig. 1(a) and (b) illustrates the conceptual diagram of the imaging principle.

454

IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 17, NO. 2, MARCH 2013

Fig. 1. Concept of the tactile sensation imaging principle. (a) Light is injected into the waveguide to totally reflect. (b) Light scatters as the waveguide deforms due to the stiff tissue inclusion.

Fig. 3. FEM model of an idealized breast tissue model. The sensing probe of TSIS is also modeled on top of the breast tissue model.

TSIS

Sensing probe Soft tissue

h d

E

Spherical inclusion

Fig. 2. Cross section of an idealized breast model for estimating inclusion parameters. The tissue inclusion has three parameters — size d, depth h, and Young’s modulus E.

III. TISSUE INCLUSION PARAMETER ESTIMATION The mechanical properties such as the stiffness and geometry such as the size are important in characterizing the embedded lesions. From the TSIS, we obtain tactile images with maximum pixel values, total pixel values, and deformation area. From these parameters we will estimate the size, depth, and elasticity of the embedded lesions. The proposed estimation method consists of the forward algorithm and an inversion algorithm. The forward algorithm is designed to predict the tactile parameters (maximum deformation, total deformation, and deformation area) based on the parameters of the tissue inclusion (size, depth, and modulus). Then these results are used in the inversion algorithm. In the inversion algorithm, we use tactile parameters obtained from the TSIS and simulated values from the forward algorithm to estimate the size, depth, and modulus of the embedded lesion. The proposed method is then validated by the realistic tissue phantoms. A. Problem Formulation To estimate tissue inclusion parameters, the forward and inversion algorithms were developed based on the idealized breast tissue model. This model is shown in Fig. 2. The assumptions

used in the model are as follows [25]. 1) The breast tissue is assumed as a slab of material of constant thickness that is fixed to a flat, incompressible chest wall. 2) The inclusion is assumed to be spherical. 3) Both tissue and inclusion are assumed linear and isotropic. 4) The interaction between the sensing probe and tissue is assumed to be frictionless. B. Forward Algorithm The purpose of forward algorithm is to find the relationship between tissue inclusion parameters and tactile data. In this paper, an FEM is considered for the forward algorithm. The FEM based on the idealized breast tissue model is performed using ANSYS ver. 11.0, an engineering simulation software package (ANSYS Inc., Pennsylvania, USA). The FEM model consists of a three-layered sensing probe, soft tissue, and inclusion. All are modeled using SOLID95 3D elements available in ANSYS. Appropriate surface-to-surface contact elements have been defined in the ANSYS database model. The model consists of 3 000 finite elements. The following assumptions are used for the FEM forward modeling. 1) The breast tissue and inclusions are elastic and isotropic. This means that the parameters of a material are identical in all directions [33], [34]. 2) Poisson’s ratio of each material is set to 0.49, because the breast tissue is elastic. This value was also used in the literature [35]. 3) The breast tissue is assumed to be on top of nondeformable hard surfaces such as bones. If the TSIS compresses against the tissue surface containing a stiff tissue inclusion, the sensing probe of TSIS deforms. In an FEM, the deformed shape of sensing probe is captured in response to different inclusion parameters, size d, depth h, and Young’s modulus E. To quantify the amount of sensing probe deformation, the following definitions are used. 1) Max1 is defined as the largest vertical disimum deformation OFEM placement of the FEM element of sensing probe from the non2 is defined as deformed position. 2) Total deformation OFEM the vertical displacement summation of FEM elements of sensing probe from the nondeformed position. 3) Deformation area 3 is defined as the projected area of the deformed surface of OFEM the sensing probe. In the forward algorithm, (d, h, E) are input

LEE AND WON: THE TACTILE SENSATION IMAGING SYSTEM FOR EMBEDDED LESION CHARACTERIZATION

455

Fig. 4. (a) Diagram of input variables (d, h, E) and output variables (O F1 E M , O F2 E M , O F3 E M ) in the forward algorithm. (b) Diagram of input variables 1 2 3 ˆˆ ˆ (O T S IS , O T S IS , O T S IS ) and output variables (d, h, E) in inversion algorithm.

TABLE I ˆ h, ˆ E) ˆ WITH RELATIVE ESTIMATION ERRORS AND DIFFERENCE ACTUAL PARAMETERS (d, h, E) AND ESTIMATED PARAMETERS (d, BETWEEN TRUE VALUE AND THE ESTIMATED VALUE OF NINE INCLUSIONS IN THE CALIBRATION TISSUE PHANTOM

1 variables and maximum deformation OFEM total deformation 2 3 OFEM and deformation area OFEM of sensing probe are output variables. The diagram of input variables (d, h, E) and output variables (OF1 E M , OF2 E M , OF3 E M ) is shown in Fig. 4(a). To investigate the relationship between the input variables 1 2 3 , OFEM , OFEM ), 134 (d, h, E) and the output variables (OFEM input variables (d, h, E) are randomly generated with the minimum and maximum constraints of d as [0 mm; 15 mm], h as [4 mm; 12 mm], and E as [40 kPa; 120 kPa] [8]. The Young’s modulus of the tissue inclusion was chosen to represent normal glandular tissue (40 kPa) and invasive ductal carcinoma (100 kPa) at 5% precompression with a loading frequency of 4.0 Hz. The Young’s modulus of the surrounding tissue was set as 5 kPa. 1 2 3 , OFEM , OFEM ) correspondThen 134 output variables (OFEM ing to 134 input variables (d, h, E) are being generated using 1 in a 3-D space, FEM. To visualize 134 output variable OFEM 1 are rescaled to [0; 255] and displayed as the values of OFEM circles at the locations specified by input variables (d, h, E). The size of each circle is determined by the values. We notice that as the size of inclusion d increases, the maximum deformation, 1 , increases as the effect of bigger tissue inclusion causes OFEM more change in the sensing probe deformation. As the depth 1 deof inclusion h increases, the maximum deformation OFEM creases as the effect of stiff inclusion gets reduced and sensing probe presses the soft tissue. Also as Young’s modulus E of 1 increases inclusion increases, the maximum deformation OFEM as the stiff inclusion makes the sensing probe to deform more. 2 and We also notice that the other two output variables OFEM 3 OFEM have the similar patterns and omit to save space.

C. Mapping Tactile Data 1 2 , OFEM , It is necessary to relate FEM tactile data (OFEM 1 2 3 and TSIS tactile data (OTSIS , OT S I S , OTSIS ). To map two different tactile data, realistic calibration tissue phantom 3 ) OFEM

with nine embedded stiff inclusions have been manufactured (CIRS, Inc., VA, USA). We custom designed the calibration tissue phantom with varying parameters (d, h, E) as shown in Table I. To map TSIS tactile data to FEM tactile data, tactile data of nine inclusions in the calibration tissue phantom were obtained using the TSIS. In order to quantify TSIS tactile data, maximum 1 2 , total pixel value OTSIS , and deformation area pixel value OTSIS 3 of pixel OTSIS of TSIS tactile data are computed. The definitions 1 2 3 , OTSIS , OTSIS ) are as follows. Maximum pixel value of (OTSIS 1 OTSIS is defined as the pixel value in the centroid of the tactile 2 is defined as the summation data 2) Total pixel value OTSIS of pixel values in the tactile data. 3) Deformation area of pixel 3 is defined as the number of pixel greater than the specific OTSIS threshold value in the tactile data. We assume that there is no noise in the tactile data. To find the relationship between TSIS tactile data to FEM 1 1 2 2 : OTSIS ), (OFEM : OTSIS ), tactile data, the graphs of (OFEM 3 3 and (OFEM : OTSIS ) were generated. Then using the linear regression method, the relationship between TSIS tactile data to FEM tactile data is found. Using three relationships, newly 1 2 3 , OTSIS , OTSIS ) can be transobtained TSIS tactile data (OTSIS 1 2 3 ). In this formed into the FEM tactile data (OFEM , OFEM , OFEM way, we relate TSIS tactile data with FEM tactile data. D. Inversion Algorithm ˆ h, ˆ E) ˆ The goal of an inversion algorithm is to estimate (d, 1 2 3 through obtained TSIS tactile data (OTSIS , OTSIS , OTSIS ). We 1 2 3 1 2 3 , OFEM , OFEM ) from (OTSIS , OTSIS , OTSIS ). estimate (OFEM ˆ ˆ ˆ Then we design an inversion algorithm to estimate (d, h, E) 1 2 3 , OFEM , OFEM ). In this paper, the using the determined (OFEM multilayered ANN is considered as an inversion algorithm [28]. The ANN is a computational model that is inspired by the structure and functional aspects of biological neural networks. The

456

IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 17, NO. 2, MARCH 2013

ANN is trained using the input variables to obtain the desired output variables. The multilayered ANN that we used consists of neurons united in layers. Each i layer is connected with i−1 and i+ 1 layers and neurons within the layer are not connected to each other. In the ANN, too many neurons would lead to overfitting and large variance of an error, but not enough neurons would cause high mean squared error results. Thus, numbers of neurons and layers were set experimentally to three layers. The first layer uses ten neurons with sigmoid activation function. The second layer uses four neurons with sigmoid activation function. The third layer uses three neurons with linear activation function. For the training ANN algorithm, the SCGA is utilized due to its simple and robustness characteristics compared to the other learning algorithm [28]. The SCGA is based upon a class of optimization techniques well known in numerical analysis as the conjugate gradient methods. The SCGA uses second-order information from the ANN but requires only O(N ) memory usage, where n is the number of weights in the ANN. To train 1 2 3 , OFEM , OFEM ) and correANN, 125 input variables (OFEM sponding output variables (d, h, E) are used. The remaining nine variables (d, h, E), which are used for the calibration tissue phantom design, are used for the validation of the proposed estimation method. IV. EXPERIMENTAL RESULTS A. Tissue Phantom Experiments To validate the proposed estimation method, tactile data of nine tissue inclusions in the calibration tissue phantom were obtained using TSIS. TSIS tactile data was quantified as 1 2 3 (OTSIS , OTSIS , OTSIS ). Then using the linear regression results, 1 2 3 , OFEM , OFEM ) are obtained. Finally, using the quanti(OFEM 1 2 3 , OFEM , OFEM ), we estimated tissue inclusion fied data, (OFEM ˆ ˆ ˆ parameters, (d, h, E). To measure the performance of the proposed estimation method, the cross-validation method called leave-one-out-cross-validation (LOOCV) metric was considered [28]. LOOCV is a special case of k-fold cross validation where k equals the number of instances in the data [28]. In this paper, after getting new TSIS tactile data of each inclusion, each inclusion parameters (d, h, E) were estimated using ANN with LOOCV. This estimation trial was performed 100 times per each inclusion and the results were averaged. The averaged estimation results are shown in Table I. The results show that the minimum and maximum relative estimation errors for the tissue inclusion size case are 0.7% and 20%. The mean estimation error is 6.99% with 6.28% standard deviation. For the depth estimation case, the minimum and maximum relative errors are 2.2% and 26.4%. The mean error is 7.17% with 7.47% standard deviation. For the Young’s modulus estimation case, the minimum and maximum relative errors are 1.1% and 11.72%. The mean error is 5.07% with 3.39% standard deviation. B. Clinical Results Clinical studies of the TSIS are currently being conducted at the Temple University Hospital, Philadelphia, PA, USA. The In-

TABLE II CLINICAL DATA FOR THREE PATIENTS

stitutional Review Board approval was obtained on 4th, March, 2011. The protocol number is 13661. Here, we present just a few pilot results. Three patients presented with a lesion that was initially detected by another modality (mammography, ultrasound, or manual palpation). When performing the TSIS scans, the doctor already knew where the lesions were located. For each lesion, ten tactile images were obtained and the estimated parameters were averaged. Table II shows estimated size and modulus of the examined breast lesions. The size truth values were provided by mammogram, and malignancy results were obtained from the pathology reports. The size of the lesion of the patient 1 was 4.86 mm and the estimated value using the proposed method was 4.44 mm. The relative estimation error is 8.58%. In the case of the patient 2, the size of the lesion was 6.73 mm and the estimated value was 6.12 mm. The relative estimation error is 9.06%. For patient 3 case, the size of the lesion was 8.08 mm and the estimated value was 7.34 mm, resulting in the 9.15% relative estimation error. Regarding the hardness estimation of the lesions, malignant breast lesion of the patient 1 had increased Young’s modulus (146 kPa), compared to benign lesions (97 kPa and 103 kPa, patient 2 and 3, respectively). The elasticity information was correlated with the malignancy data from the pathology reports. The stiffer lesion was malignant as shown in Table II V. DISCUSSIONS The tissue inclusion parameter estimation method, proposed in this paper, is a phantom or patient dependent approach. The forward algorithm using the FEM has been constructed based on the idealized breast tissue model. Different women have different breast thickness, shape, and Young’s modulus so the parameters of the FEM model should be updated for the better estimation results. For this purpose, FEM model is parameterized and can be easily changed depending on the different patient geometry. To map TSIS tactile data to FEM tactile data, only nine breast phantom inclusions were used. Though we showed that the relationship between TSIS tactile data and FEM tactile data can be approximately linear, the number of sample data is not enough to describe the whole relationship between TSIS tactile data and FEM tactile data. If we use more tissue inclusions for the mapping purpose, the relationship would be more accurate and estimation errors will be reduced. Generally, the malignant tumor does not have a well circumscribed round shape that the assumptions we used in the FEM. In fact, the morphological shape is the basis of mass and tumor distinction in many imaging techniques. In this paper, however, we assume that the tumor is the spherical round shape to simplify the FEM calculation. Also we use 0.49 as the Poisson’s

LEE AND WON: THE TACTILE SENSATION IMAGING SYSTEM FOR EMBEDDED LESION CHARACTERIZATION

ratio with the single breast tissue layer. This assumption however, cannot be true in the case of real breast cancer patient depending on the patient tissue characteristics. Thus, we use Poisson ration as 0.49 which is normally used in the tissue modeling. However, each patient havs the different Poisson’s ratio, in the real clinical trials, to reduce the final estimation error, the forward and inversion algorithms should be iteratively calculated. Conventionally, Young’s modulus can be calculated from the tensile stress and the tensile strain. In this paper, however, the authors found the Young’s modulus from the computational method, forward and inversion algorithms. This is the indirect method to find the Young’s modulus without tensile stress and tensile strain values. Thus, the proposed method does not require the force or stress information to calculate the Young’s modulus. The measured tissue inclusion must be smaller than the TSIS sensing probe area. The common breast tumor size is approximately 20 mm or smaller in stages 0 and I. In stage II, it is approximately 20 mm to 50 mm [38]. To detect the 50-mm breast tumor, the sensing probe should be bigger than 50 mm. The parameter estimation problem can be considered as the circled loop problem. In this paper, the forward and inversion approach had been taken as a one loop. The initial forward modeling has been done using the average breast cancer patient biological model. In the real clinical situation, however, the patient breast geometry and their parameters can be different patient by patient. Thus, the forward and inversion approach should be performed iteratively until the final estimation errors are acceptable. There are large variations in the tissue inclusion parameter estimation errors. The one reason of this is because of the variation of the ANN performance. In this paper, we utilized the SCGA as the ANN learning algorithm for its simplicity and low complexity. Despite its advantages, however, we found that the SCGA fails to find the final optimal solution of ANN in some cases. In future works, we will consider to use advanced ANN learning algorithm such as Levenberg-Marquardt algorithm (LMA). The LMA interpolates between the Gauss–Newton algorithm (GNA) and the gradient descent method. The LMA is more robust than the GNA, which means that in many cases it finds an optimal solution even if it starts very far off the final minimum. We believe that if we use the advanced ANN learning algorithm, the final parameter estimation error will be decreased. In the phantom experiment, the indentation force had been controlled by the external machine. In the clinical experiment, if we assume that the sensed material is homogeneous and isotropic, the relationship between the normal force and the integrated pixel value of the tactile image will be linear. Thus, the applied force can be estimated from the integrated pixel values using the normal force versus the integrated pixel values table, which is previously obtained by the calibration. Then the applied stress, which is the force per unit area, is obtained by dividing the applied force by the contact area. However, if the sensed material is inhomogeneous and anisotropic, the indentation force estimation using the calibration curve will contain the errors. The other way is to attach the additional force sensor on the TSIS probe and measure the indentation force directly. These

457

two approaches for the indentation force estimation without external machine will be verified in the future clinical studies. Regarding the depth information of the lesion, the accurate ground truth value should be obtained using CT or MRI that we have not included in this paper. Thus, full parameters such as size, depth, and Young’s modulus of the inclusion have been estimated only through the realistic phantom studies and validated. In the small clinical trials, only the size, and Young’s modulus estimation have been performed. For the future clinical trials, the full parameters such as size, depth, and Young’s modulus estimation trials will be performed under the large number of patient set. In the forward algorithm, the healthy surround tissue around the lesion has been set as the normal breast tissue Poisson’s ratio. However, as the reviewer pointed out, in the real clinical situation, the depth estimation error can be increased, as the depth is increasing or the modeling parameter of the healthy tissue that we set is far from the patient’s breast tissue elasticity. To prevent this, the iterative procedure between forward and inversion algorithm should be considered to perform to reduce the final estimation error. The iterative forward and inversion algorithms for the real clinical trials will be designed as the future work. The patient-specific factors such as breast density ratio between patients are very important. In the following future clinical trials, the intra patient comparison such as breast density will also be compared. VI. CONCLUSION In this paper, a tissue inclusion parameter estimation method based on tactile images proposed to quantify absolute stiffness and geometric parameters of tissue inclusions. The mechanical properties estimation is performed based on the tactile data obtained by the TSIS. In the estimation method, we used FEMbased forward algorithm and ANN-based inversion algorithm. The performance of the method was experimentally verified using realistic tissue phantoms with embedded inclusions to represent tumors. The experimental results showed that the proposed estimation method estimated the size, depth, and Young’s modulus of a tissue inclusion with 6.98%, 7.17%, and 5.07% relative error, respectively. The small-scale clinical trial results showed the potential to discern malignant lesions from benign ones using TSIS. This study is the initial step toward achieving a TSIS and associated parameter estimation method for early breast tumor detection and characterization. REFERENCES [1] F. Kamangar, G. M. Dores, and W. F. Anderson, “Patterns of cancer incidence, mortality, and prevalence across five continents: Defining priorities to reduce cancer disparities in different geographic regions of the world,” J. Clin. Oncol., vol. 24, pp. 2137–2150, 2006. [2] A. Jemal, R. Siegel, E. Ward, Y. Hao, J. Xu, and M. J. Thun, “Cancer statistics, 2009,” Cancer J. Clin., vol. 59, pp. 225–249, 2009. [3] H. Shojaku, H. Seto, H. Iwai, S. Kitazawa, W. Fukushima, and K. Saito, “Detection of incidental breast tumors by noncontrast spiral computed tomography of the chest,” Radiat. Med., vol. 26, no. 6, pp. 362–367, 2008. [4] A. Fenster and D. B. Downey, “3-D ultrasound imaging: A review,” IEEE Eng. Med. Biol. Mag., vol. 15, no. 6, pp. 41–51, Nov./Dec. 1996.

458

IEEE JOURNAL OF BIOMEDICAL AND HEALTH INFORMATICS, VOL. 17, NO. 2, MARCH 2013

[5] H. Degani, V. Gusis, D. Weinstein, S. Fields, and S. Strano, “Mapping pathophysiological features of breast tumors by MRI at high spatial resolution,” Nat. Med., vol. 3, pp. 780–782, 1997. [6] C. R. Gentle, “Mammobarography: A possible method of mass breast screening,” J. Biomed. Eng., vol. 10, no. 2, pp. 124–126, 1988. [7] D. B. Kopans, “Clinical breast examination for detecting breast cancer,” J. Amer. Med. Assoc., vol. 283, no. 13, pp. 1270–1280, 2000. [8] T. A. Krouskop, T. M. Wheeler, F. Kallel, B. S. Garra, and T. Hall, “Elastic moduli of breast and prostate tissues under compression,” Ultrason. Imag., vol. 20, no. 4, pp. 260–274, 1998. [9] T. Ratanachaikanont, “Clinical breast examination and its relevance to diagnosis of palpable breast lesion,” J. Med. Assoc. Thai., vol. 88, pp. 505– 507, 2005. [10] J. Ophir, I. Cespedes, H. Pennekanti, Y. Yazdi, and X. Li, “Elastography, a quantitative method for imaging the elasticity of biological tissues,” Ultrason. Imag., vol. 13, no. 2, pp. 111–134, 1991. [11] L. Gao, K. J. Parker, R. Lerner, and S. Levinson, “Imaging of elastic parameters of tissue—A review,” Ultrasound Med. Biol., vol. 22, no. 8, pp. 959–977, 1996. [12] J. Rogowska, N. A. Patel, J. G. Fujimoto, and M. E. Brezinski, “Optical coherence tomographic elastography technique for measuring deformation and strain of atherosclerotic tissues,” Heart, vol. 90, no. 5, pp. 556– 562, 2004. [13] A. Vinckier and G. Semenza, “Measuring elasticity of biological materials by atomic force microscopy,” FEBS Lett., vol. 430, no. 1–2, pp. 12–16, 1998. [14] H. Rivaz, E. Boctor, P. Foroughi, R. Zellars, G. Fichtinger, and G. Hager, “Ultrasound elastography: A dynamic programming approach,” IEEE Trans. Med. Imag., vol. 27, no. 10, pp. 1373–1377, Oct. 2008. [15] A. Stravros, D. Thickman, M. Dennis, S. Parker, and G. Sisney, “Solid breast nodules: Use of sonography to distinguish between benign and malignant lesions,” Radiology, vol. 196, no. 1, pp. 79–86, 1995. [16] K. S. Bhatia, D. D. Rasalkar, Y. P. Lee, K. T. Wong, A. D. King, Y. H. Yuen, and A. T. Ahuja, “Real-time qualitative ultrasound elastography of miscellaneous non-nodal neck masses: applications and limitations,” Ultrasound Med. Biol., vol. 36, no. 10, pp. 1644–1652, 2010. [17] M. F. Insana, C. Pellot-Barakat, and M. Sridhar, K. K. Lindfors, “Viscoelastic imaging of breast tumor microenvironment with ultrasound,” J. Mammary Gland Biol. Neoplasia., vol. 9, no. 4, pp. 393–404, 2004. [18] J. Dargahi and S. Najarian, “Human tactile perception as a standard for artificial tactile sensing-A review,” Int. J. Med. Robot. Comput. Assist. Surg., vol. 1, no. 1, pp. 23–35, 2004. [19] M. Eltaib and J. Hewit, “Tactile sensing technology for minimal access surgery—A review,” Mechatronics, vol. 13, no. 10, pp. 1163–1177, 2003. [20] A. Mojra, S. Najarian, S. M. TowliatKashani, F. Panahi, and M. Yaghmaei, “A novel haptic robotic viscogram for characterizing viscoelastic behavior of breast tissue in clinical examinations,” Int. J. Med. Robot. Comput. Assist. Surg., vol. 7, no. 3, pp. 282–292, 2011. [21] S. M. Hosseini, S. M. Kashani, S. Najarian, F. Panahi, S. M. Naeini, and A. Mojra, “A medical tactile sensing instrument for detecting embedded objects with specific application for breast examination,” Int. J. Med. Robot. Comput. Assist. Surg., vol. 6, no. 1, pp. 73–82, 2010. [22] P. Dario and M. Bergamasco, “An advanced robot system for automated diagnostic tasks through palpation,” IEEE Trans. Biomed. Eng., vol. 35, no. 2, pp. 118–126, Feb. 1998. [23] R. D. Howe, W. J. Peine, D. A. Kantarinis, and J. S. Son, “Remote palpation technology,” IEEE Eng. Med. Biol., vol. 14, no. 3, pp. 318–323, May/Jun. 1995. [24] V. Egorov and A. P. Savazyan, “Mechanical imaging of the breast,” IEEE Trans. Med. Imag., vol. 27, no. 9, pp. 1275–1287, Sep. 2008. [25] P. S. Wellman, E. P. Dalton, D. Krag, K. A. Kern, and R. D. Howe, “Tactile imaging of breast masses: First clinical report,” Arch. Surg., vol. 136, no. 2, pp. 204–208, 2001. [26] G. Weber, “Using Tactile Images to Differentiate Breast Cancer Types,” Ph.D. dissertation, Division of Engineering and Applied Sciences., Harvard University, Cambridge, MA, USA, 2000 [27] A. M. Galea, “Mapping tactile imaging information: Parameter estimation and deformable registration,” Ph.D. dissertation, School of Engineering and Applied Sciences, Harvard University, Cambridge, MA, USA, 2004 [28] C. M. Bishop, Pattern recognition and machine learning, (Series Information Science and Statistics). New York, USA: Springer, 2007.

[29] J.-H. Lee and C.-H. Won, “Inclusion mechanical property estimation using tactile images, finite element method, and artificial neural network,” in Proc. IEEE 33th Annu. Int. Conf. IEEE Eng. Med. Biol. Soc., Boston, MA, USA, Aug. 2011, pp. 14–17. [30] G. S. Rajan, G. S. Sur, J. E. Mark, D. W. Schaefer, and G. Beaucage, “Preparation and characterization of some unusually transparent poly (dimethylsiloxane) nanocomposites,” J. Polymer. Sci. B Polymer Phys., vol. 41, pp. 1987–1901, 2003. [31] C. O. Flynn and B. A. McCormack, “A three-layer model of skin and its application in simulating wrinkling,” Comput. Methods. Biomech. Biomed. Eng., vol. 12, no. 2, pp. 125–134, 2009. [32] G. Keiser, Optical Fiber Communications, 3rd ed. New York, USA: McGraw-Hill, 1999. [33] Y. C. Fung, Biomechanics: Mechanical Parameters of Living Tissues. New York, USA: Springer-Verlag, 1993. [34] K. J. Parker, S. R. Huang, R. A. Musulin, and R. M. Lerner, “Tissue response to mechanical vibrations for sonoelasticity imaging,” Ultrasound Med. Biol., vol. 16, no. 3, pp. 241–246, 1990. [35] A. Keller, T. R. Nelson, L. I. Cervino, and J. M. Boone, “Simulation of mechanical compression of breast tissue,” IEEE Trans. Biomed. Eng.,, vol. 54, no. 10, pp. 1885–1891, Oct. 2007. [36] R. Picard and D. Cook, “Cross-validation of regression models,” J. Amer. Stat. Assoc., vol. 79, no. 387, pp. 575–583, 1984. [37] A. L. Kellner, T. R. Nelson, L. I. Cervino, and J. M. Boone, “Simulation of mechanical compression of breast tissue,” IEEE Trans. Biomed. Eng., vol. 54, no. 10, pp. 1885–1891, Oct. 2007. [38] S. G. Orel and M. D. Schnall, “MR imaging of the breast for the detection, diagnosis, and staging of breast cancer,” Radiology, vol. 220, no. 1, pp. 13– 30, 2001.

Jong-Ha Lee received the B.S. degree in electronics engineering in 2000 from Inha University, Incheon, Korea, the M.S. degree in electrical engineering in 2005 from Polytechnic Institute of New York University, Brooklyn, New York, USA, and the Ph.D. degree in electrical engineering from Temple University, Philadelphia, PA, USA. He was with Samsung advanced institute of Technology as a Research Staff Member. He is currently an Assistant Professor with the Department of Biomedical Engineering, School of Medicine, Keimyung University, Daegu, Korea. His current research interests include tactile sensation imaging for tissue characterization, computer-aided diagnosis, medical image analysis, pattern recognition, and machine learning.

Chang-Hee Won received the B.S., M.S., and Ph.D. degrees in electrical engineering from the University of Notre Dame, Notre Dame. IN, USA. He was a Senior Research Engineer with the Department of Satellite Communications Systems, Electronics and Telecommunications Research Institute, University of Notre Dame. He is currently an Associate Professor in the Department of Electrical and Computer Engineering and Director of Control, Sensor, Network, and Perception Laboratory at Temple University, Philadelphia, PA, USA. His research interests include sensors, image processing, and stochastic optimal control. In the applied research, his interests include tactile imaging sensor for biomedical applications.