Semi-automated experimental setup integrated with a system-of-biochips (SoB)

The experimental setup consists of the schematic shown in Fig. 1C which is a semi-automated system with two horizontal probes (X-direction) and one vertical probe (Z-direction) for interacting and taking measurements from the tissue sample. The system integrates three biochips, namely S1, S2, and S3 (as shown in Fig. 1C inset), which constitutes the system of biochips (SoB). S1 acts as the microforce sensor for mechanical characterization, while S2 and S3 are used for electro-thermal characterization. The photograph of the fabricated biochips is shown in Fig. 2A. The design of S1 consists of four piezoresistive bridges fabricated on a thin diaphragm created on a silicon substrate. The dimension of S1 is 2.5 mm x 2.5 mm. The optical microscope image of one piezoresistive bridge is shown in Fig. 2B. The functional elements of S2 and S3 consist of a microheater, interdigitated electrodes, and thermistors around the microheater. While the microheater is used to heat the tissue, the interdigitated electrodes and the thermistors measure the electrical and thermal properties of the tissue placed between S2 and S3. A square shape profile is provided for the microheater, thermistors, and the electrodes to geometrically match with the cuboidal shape of the sample tissues, which makes subsequent data analysis and parameter extraction simpler. The biochip has an overall dimension of 7 mm x 12 mm with an active area of 1 mm x 1 mm. An optical profilometry image of the sensing elements is shown in Fig. 2C.

Fig. 2

System-of-biochips (SoB) for electro-thermo-mechanical phenotyping of breast tissues. A optical photograph of the biochips, B optical microscope image showing the piezoresistive bridge in the force sensor for mechanical characterization, and C optical profilometry image of the electro-thermal biochip showing the microheater, interdigitated electrodes and thermistor

Fig. 3

Experimental setup. A Photograph of the semi-automated system with the three probes and system of biochips, B close-up view of one electrothermal biochip and the force sensing chip connected to the probes, and C experimental protocol for the ETM measurements (i) sample is loaded into the system, (ii) S1 indents the tissue followed by stress-relaxation for 150s, (iii) S1 retracts, S2 and S3 probe the tissue, make contact, and heats the sample to 37 ℃, and (iv) measurements are captured from the RTDs followed by the impedance characterization

The photograph of the assembled system with the probes, electronic modules, biochips, and rotary platform for placing the sample is shown in Fig. 3A. The biochip S1, which integrates the force sensor, is wire bonded on a carrier printed circuit board and attached to the vertical probe along with a force transfer mesa, while the biochips with the electro-thermal sensors, S2 and S3, are attached to the two horizontal probes. This arrangement is shown schematically in the inset of Fig. 1C, and the actual photograph of the arrangement is shown in Fig. 3B. The probes are actuated through NEMA-17 stepper motors connected to a microcontroller and motor driver system. The biochips are connected to the electronic modules in the system using flexible flat cables (FFC). The electronic modules integrate the circuits for multiplexing and measuring the voltages from the force sensor and the resistances from the thermistors and interdigitated electrodes. The microheater on S2 is connected to a constant current voltage driver circuit to act as the heat source for the thermal characterization of the tissue. The microheater structure on S3 acts as a thermistor to detect the heat transmitted through the tissue, along with the other three thermistors around it. Finally, the interdigitated electrodes on S2 and S3 are connected externally to the GW-INSTEK LCR 8105G impedance analyzer system for electrical measurements. The complete actuations and measurements are controlled through a laptop connected to the system through the microcontroller via the UART port through serial communication. Before loading the samples, a disposable 3D printed attachment for holding the tissue surface is placed on the rotary platform to avoid contamination of the setup (shown in Fig. 3B).

Protocol for electro-thermo-mechanical characterization of breast biopsy tissues

The experiments are conducted in a class 10000 clean room with a controlled ambient temperature of 21 ℃. The experimental methodology of evaluating the system involves capturing the electrical, thermal, and mechanical properties of the tissue sample loaded into the system using the SoB. The Z-axis probe indents the sample vertically and captures the mechanical force response and stress relaxation using the microforce sensor attached to it. The mesa interacts with the tissue and transfers the force to the diaphragm in the sensor S1, and the changes in the piezoresistive bridges are recorded by the onboard electronics through the wires attached to the carrier PCB. The two X-axis probes capture the electro-thermal properties of the sample, such as its electrical impedance and thermal conductivity, using the biochips attached to them. The microheater on S2 heats the tissue to the required temperature, and the interdigitated electrodes between S2 and S3 capture the electrical impedance data across the tissue sample. This methodology is chosen so that once the sample is loaded into the system, its complete biophysical phenotype (electrical, thermal, and mechanical) can be captured. The mechanical characterization, which involves loading the sample to 30 mN force followed by stress-relaxation and unloading to the original state, is performed first as the electro-thermal characterization involves heating the tissue sample to 37 ℃. Once the sample is loaded into the system, the following protocol is used to capture the ETM properties (also shown graphically in Fig. 3C (i to iv):

1.

The Z-axis probe with the force sensor indents the sample at a constant rate of 30 μm/s to a compressive load of 30 mN. The sample is kept loaded for 150 s to capture the relaxation data, and then the Z-axis probe is moved up to unload the sample. The load-displacement data is captured from the force sensor.

2.

After this, the two X-axis probes, with the biochips for electro-thermal characterization, approach the sample at 20 μm/s and make contact.

3.

The microheater on S2 is switched on and is used to heat the tissue to a temperature of 37 ℃ in steps of 3 ℃. At each temperature point, the resistance values of the thermistors are recorded. These resistance values are then mapped to the sensed temperature and used to compute the thermal conductivity.

4.

At 37 ℃, the impedance of the sample from 10 Hz to 3 MHz at 100 mV excitation voltage is captured across the IDEs on the two biochips.

5.

After the impedance and thermistor measurements at 37 ℃ are completed, the X-axis probes move away from the sample, thereby completing the measurements.

A representative video showing the system in operation is provided as a Supplementary Video. To clearly indicate the clearances between the probes, this video shows a position where all probes are in contact with the tissue. For the actual experiments, the Z-axis probe indents the tissue and moves back, followed by the measurements using the X-axis probes (as detailed in Fig. 3C). The video captured with a larger-sized tissue for clear visibility of the different components is shown.

Experimental results and biophysical parameters from the ETM characterization

The electrical impedance at 37 ℃ as a function of frequency, the conducted thermal energy through the tissue, and the mechanical loading and relaxation response are the electrical (E), thermal (T), and mechanical (M) measurements, respectively, performed on each sample. Figure 4 shows the summary of the ETM characterization measurements of the samples from N = 14 subjects indicating the mean curves with error bands for the three sample groups, viz. adjacent normal (AN), fibroadenoma (FA), and carcinoma (CA). Since the experiment protocol involves heating the tissue, which might cause irreversible changes, measurements were not repeated on the same sample. However, for each of the 10 carcinoma, 4 fibroadenoma, and 14 adjacent normal tissue samples, two samples each were extracted and measured. Thus, experiments were conducted from a total of n = 56 samples from N = 14 subjects. Additionally, to quantify the variations between experiments, the electrical impedance and mechanical characterization experiments without heating the tissue were carried out on a few samples multiple times (n = 3). The maximum coefficient of variation observed was 6.5%.

Fig. 4

Mean plots of ETM measurements from N = 14 paired breast biopsy samples. A and B shows the magnitude of the complex impedance and imaginary component of the complex impedance as a function of frequency, C the mechanical loading characteristics obtained for the samples, D viscoelastic relaxation curves and E the thermal response of the samples to heating showing the temperature at source and sense points

Figure 4 A and B show the mean plots of the magnitude and imaginary part of the complex electrical impedance as a function of frequency measured at a tissue temperature of 37 ℃. The impedance magnitude plots of each of the 14 sample pairs are shown in Fig. S1. The plot of the mean phase response and the plot of the real part of the impedance vs. the imaginary part for the AN, CA, and FA groups is also shown in Fig. S2. The mean impedance of CA is observed to be higher than AN and FA at all the measured frequencies. The mean plot of the imaginary part of the impedance (Fig. 4B) has a continuously decreasing trend for the AN group, while for the FA and CA samples, there is a kink in the low-frequency region. However, the trend of higher mean impedance for the CA group compared to the AN and FA is also observed for the imaginary part.

The impedance curve data for the sample groups also facilitates the extraction of circuit parameters which can provide insights into the tissue organization. The standard cole-cole model is modified with two capacitances, Cdl, to account for the double layer capacitance at the tissue-electrode interface at both ends of the sample. The other circuit parameters of the tissue are similar to the Cole-Cole model, namely, the extracellular resistance Re, the membrane resistance Rm, the membrane capacitance Cm, and the intra-cellular resistance Ri. The equivalent circuit used to fit the mean impedance curves is shown in Fig. S3. The fitted curves for the mean impedance and the mean phase response along with the experimental data for the three groups, are shown in Figs. S4 and S5, respectively. The CA samples had significantly (p < 0.001) higher Re (1.75e5 ± 7.07e3 Ω) and Ri (3.29e4 ± 1.49e3 Ω) than FA (Re = 3.51e4 ± 4.34e3 Ω; Ri = 1.3e4 ± 2.33e2 Ω) and AN (Re = 2.27e4 ± 1.78e3 Ω; Ri = 1.23e4 ± 5.31e2 Ω). Rm was calculated to be highest for the FA (5.48e6 ± 1.14e5 Ω, p < 1e-5) samples and similar for CA (6.43e4 ± 3.14e3 Ω) and AN (9.27e4 ± 1.33e4 Ω). The membrane capacitance, Cm was calculated to be highest for the CA (1.02e-10 ± 1.77e-11 F, p < 0.05) samples while it was similar for the FA (1.11e-11 ± 6.94e-13 F) and AN (1.1e-11 ± 1.63e-12 F) groups. All the fitted circuit parameter values are summarized in Table S1.

The mechanical loading and stress relaxation tests plots for the three groups are shown in Fig. 4C and D, respectively. Figure 4C shows the changes in the indentation force measured by the force sensor, S1, with the indentation depth of the tissue, with the loading force capped to a maximum of 30 mN. The error bars indicate the family of loading curves for each sample within the sample groups. The data for each sample pair is provided in Fig. S6. It can be seen that the mean loading curves for the CA and FA samples overlap each other and are steeper than the AN sample, indicating higher stiffness. The AN samples get indented to a larger extent (larger amount of strain) for the same applied force of the indenter. Figure 4D plots the relaxation in the normalized load with time up to 150 s. It can be seen that the FA group has the maximum relaxation, followed by CA and AN. While in the loading curves, the CA and FA groups were observed to behave in a similar manner, the relaxation curves revealed differences between the two groups, suggesting the utility of such a characterization.

Figure 4E summarizes the results of the thermal characterization of the samples. The x-axis plots the temperature at the source point of heating of the samples using the on-chip microheater in S2. The samples were heated from room temperature to 37 ℃. Given that in-vitro collagen is thermally unstable beyond 37 ℃ [19], the samples were not heated beyond this temperature to avoid any irreversible damage or charring of the tissue that could confound the impedance measurements at 37 ℃. The y-axis plots the temperature detected by the thermistors in S3 from the heat transmitted through the tissue. It can be seen that the AN samples attain a higher temperature at the sense point for a given source temperature than the CA and FA groups, indicating higher thermal conductivity. The CA and FA groups attain similar temperatures at the sense point for the given source temperature, with the CA group having a comparatively lower value than the FA group.

Fig. 5

Biophysical parameters extracted from the ETM measurements. A impedance at 15 kHz and 37 ℃, B thermal conductivity, K, C mechanical stiffness, k, and D percentage relaxation in load

From the experimental data in Fig. 4, four key biophysical parameters were extracted. These are namely, the impedance \(Z\) (Ω) at 15 kHz, the thermal conductivity \(K\) (Wm-1 K-1) computed at 37 ℃, the mechanical stiffness \(k\) (kN/m), and the percentage relaxation in normalized load, \(\%R\) (in %). These extracted parameters for each of the three sample groups are shown as scatter interval plots in Fig. 5. All the scatter plots were first assessed for their conformity to the normal distribution using the Shapiro-Wilk test and were observed to follow the distribution. The frequency of 15 kHz was chosen as the non-zero frequency, which showed the highest magnitude of difference between the three groups in Fig. 4A. The thermal conductivity, \(K\) of the samples, were computed at 37 ℃ from the mean resistance measured by the thermistors, power input provided to the microheater, and the geometry of the samples. The stiffness, \(k\) was calculated as the slope of the loading curves for each sample at 20% strain. The percentage relaxation, \(\%R\) value was selected as the endpoint relaxation at 150 s of the stress relaxation experiments.

From Fig. 5A, the CA group had a significantly higher mean impedance of 110018.8 ± 20293.8 Ω than FA (44261.8 ± 10496.2 Ω, p = 2.02e-2) and AN (22206.2 ± 3400.7 Ω, p = 2.6e-3). The impedances of the FA and AN groups were not found to be significantly different from each other. The AN group had a significantly higher mean thermal conductivity of 0.45 ± 0.025 Wm-1 K-1 compared to the FA (0.242 ± 0.019 Wm-1 K-1, p = 6.4e-5) and CA (0.189 ± 0.018 Wm-1 K-1, p = 7.64e-8) groups (Fig. 5B). The \(K\) of FA and CA were not significantly different. The CA samples had the highest mean stiffness of 0.076 ± 0.009 kN/m, which was significantly higher (p = 1.4e-4) than the AN (0.02 ± 0.003 kN/m) (Fig. 5C). The FA samples were also significantly stiffer than the AN with \(k\) of 0.057 ± 0.003 kN/m (p = 1.6e-4), which was statistically similar to the CA group. Finally, the FA samples had the highest relaxation in normalized load among the three groups of 47.8 ± 5.1% compared to CA (37.2 ± 2.5%) and AN (26.4 ± 1.75%) (Fig. 5D). The percentage relaxation observed for the FA and CA groups was significantly higher than in the AN sample (p < 0.01). The extracted biophysical parameter values from each of the 14 sample pairs are summarized in Table 1.

Table 1 Extracted biophysical parameters from N=14 paired samplesAnalysis with a combination of the biophysical parameters

As is evident from Fig. 5, no single biophysical parameter extracted from the experimental data is able to differentiate between all the three sample groups (AN, FA, and CA) with statistical significance. The electrical impedance at 15 kHz (Z) is able to differentiate CA from AN and FA, but not AN from FA. Likewise, the thermal conductivity (K), stiffness (k), and percentage relaxation in load (%R) are able to differentiate AN from FA and CA, but not FA from CA. Fisher’s combined probability test was applied to understand whether an analysis with a combination of parameters could differentiate between all the three groups. Before applying the test, each parameter was pairwise tested for its independence to assess whether they add new information that enables differentiation with higher statistical significance.

The pairwise testing was performed by plotting each sample on a 2D space defined by two of the four parameters and then repeated for each parameter pair. The independence was assessed by calculating the R-square value of the linear fit for each sample group. A low R-square indicates a poor fit and thereby a higher degree of independence. The pairwise scatter plots are shown in Fig. 6A–F. It can be seen from the analysis that Z - K and K - %R are the most independent pair of parameters, with the maximum R-square value of only 0.09 for the AN group in the K - %R plot. The k - %R pair was noted to be the least independent pair of parameters with R-square values of 0.48, 0.63, and 0.008 for the AN, FA, and CA samples. The lower degree of independence for the k - %R pair follows from the fact that both parameters are extracted from different aspects of the mechanical characterization of the samples. All the plots also show a clear separation between the AN and CA groups, with the FA group clustering between the two in a partially overlapped manner. A colormap of the R-square values for each of the three groups is shown in Fig. 6G–I. The colormaps indicate that the parameters are least independent for the FA samples, suggesting difficulty differentiating it from AN and CA, as was also evident from the experimental data. The Fisher’s combined probability test with all the four parameters in a multi-modal manner was able to differentiate between the three groups with the highest statistical significance. The test differentiated AN from FA, AN from CA, and CA from FA with p = 1.68e-9, p = 5.76e-13, and p = 4.5e-3 respectively. Thep values for the tests with the other combinations of the parameters, including those for the single parameter Student t-tests for the three group comparisons, are summarized in Table 2. Since only Z was able to differentiate FA from CA and K, k, and %R was able to distinguish between AN and FA and AN and CA, combining Z with one of K, k, or %R was able to differentiate between all the three groups, albeit with lower statistical power. Additionally, though K or k alone was not able to distinguish CA from FA, testing with the combination of these two parameters was able to achieve the differentiation of CA from FA with a low degree of statistical significance. The data in the table, in summary, shows that since the parameters are fairly independent and add information to the differentiation, the p values progressively become lower as more parameters among the four are used as a basis for differentiation between the sample groups.

Table 2 Analysis of combination of modalities Fig. 6

Evaluation of independence of the modalities. A–F shows the 2D scatter plot of the pair-wise independence plots between the four modalities, namely, impedance at 37 ℃ (Z), thermal conductivity (K), stiffness (k), and percentage relaxation in normalized load (%R), and (G), (H) and (I) show the heatmap of the R-square values indicating the level of independence between the modalities for each class of samples, namely, adjacent normal (AN), fibroadenoma (FA), and carcinoma (CA)

Evaluation of classification accuracy using a gaussian process classifier

While Fisher’s combined p-value analysis only gives insights into the potential of the combination of parameters to better differentiate between the sample groups, it does not however perform any classification or assess its accuracy. For this, Gaussian process classifiers with different covariance kernels, namely, rational quadratic, squared exponential, exponential, and matern 5/2 were evaluated with the four biophysical parameters (Z, K, k, and %R) as input features to classify the samples into the three sample groups (AN, FA, and CA). The classification was performed using the leave-one-out cross-validation (LOOC) technique. The matern 5/2 kernel was overall found to be the best performing in terms of the RMSE. Figure 7 summarizes the key results from the Gaussian process classification using the matern 5/2 covariance kernel. The four parameters (Z, K, k, and %R) standalone and in various combinations were used as input features for the classification. Figure 7A–D shows the classification response plots along with the RMSE when each of the parameters Z, K, k, and %R alone was used as the metric for classification. Using Z as the feature gives the poorest RMSE of 0.6932. Using only K or k gives a better RMSE of 0.4442 and 0.4286, respectively, as compared to Z (0.6932) and %R (0.5914). However, the FA group has been significantly misclassified or poorly classified in all four cases, contributing significantly to the higher RMSE. The lowest RMSE of 0.2419 is obtained when all the four parameters are used as input features for the classification task (Fig. 7E). Using all the four parameters was also able to bring down the error in the classification of the FA group. Table S2 summarizes the RMSE values obtained for all the combinations of the parameters for the four different covariance kernels used for the classification. These results further corroborate the highly significant p-values obtained using the Fisher’s combined p-value tests with all the parameters.

Fig. 7

Evaluation of prediction errors with the gaussian process classifier using (A) only Z, (B) only K, (C) only k, (D) only percentage relaxation (%R), and (E) all four parameters for training the classifier