Optimizing Design With Extensive Simulation Data: A Case Study of Designing a Vacuum-Assisted Biopsy Tool

Computational simulation is increasingly used to assist the engineering design process. However, current software tools provided for working with simulations do not complement the human creative design process. Two deficiencies observed in current simulation tools are (1) a lack of ability of designers to understand, compare, and make decisions based on alternatives in a large set of related simulations, and (2) the requirement of focus on low-level details, such as boundary conditions, that can distract designers from their goals. These details may only be understood and guided by an expert in computational methods.

Design by Dragging (dbd) is an interactive software tool proposed to address the aforementioned major obstacles [1]. Extending the previous study, this paper presents an advanced case study with higher simulation complexity and a larger design space to explore optimal designs. The evaluated space was a virtual design of a vacuum-assisted biopsy (VAB) tool. In this section, a systematic review is provided to facilitate the understanding of the current clinical issues of VAB and the available scientific data that can be used to develop a simulation model.

Vacuum-assisted biopsy technology is designed to provide larger core tissue samples for breast cancer diagnosis than other currently available needle biopsy technologies. By using a rotational cutting method [2], this technology can remove tough tissues such as calcifications. However, the technology has shortcomings and limitations. Underestimation rates of breast cancer diagnoses from VAB samples have been reported in retrospective studies conducted by medical institutions [3–6]. The results of these studies suggest that underestimation rates were 8.1% and 13.0% for an 8- and 11-gauge needle, respectively. In addition, a few studies have reported nonsignificant sample sizes and blank sample collections [7,8]. In some of these studies, the VAB cutting mechanism was not sufficiently strong to traverse through the dense tissue. In other studies, tiny tissue samples were obtained when sampling soft tissues such as the adipose tissue. Occasionally, a “dry tap” occurred, indicating that the tissue sample was not fully separated from the main tissue. In such cases, zero tissue volume is retrieved in a sampling sequence, and the patient must undergo rebiopsy.

The ability to simulate the VAB tissue-sampling process is critical for evaluating the design of the VAB tool. In particular, the tissue cutting force during the retrieval process is a crucial indicator of the tool's cutting ability. On the basis of this information, a designer can select optimal cutting speed configurations and appropriate input power sources while maintaining strong tool ergonomics and reasonable component costs.

The rotational cutting method of VAB involves a hollow cylindrical inner cutter being driven to simultaneously rotate and translate to cut the tissue. The method applies force to the tissue surface in normal and tangential directions. The cutting behavior is similar to slicing by using a knife or a line cutter [9–11]. The introduction of the tangential force is the key to improve tissue cutting performance. The ratio of the tangential component of the cutting speed to the normal component of the cutting speed, called “slice-push ratio,” is the main factor affecting the cutting force [12]. With a fixed translational cutting speed, increasing the slice-push ratio can significantly reduce the cutting force.

In order for developing a computational model of the needle cutting soft tissue, material models that describe the mechanical deformation of the three main types of breast tissue (adipose, fibroglandular, and tumor) are found in the literature [13–17]. Hyperelastic models for adipose and fibroglandular tissues are available in neo-Hookean and polynomial forms, and Young's modulus is approximated for tumor tissue.

In this study, a simulation process was developed based on the information collected from the literature. Critical factors such as the slice-push ratio, tissue type, and translational cutting speed were specified as the simulation input parameters. The mechanical behavior of the breast tissue during tissue sampling was predicted, including stress distribution, reaction force on the cutter, and tissue damage. The results were subsequently used to calculate the tool's performance attributes. The developed process enabled numerous input parameter configurations to be simulated to populate a large design space, which was subsequently imported into the dbd system. Three cases were presented to show how the user can intuitively interact with the design space in the dbd environment and seek out optimal design solutions. The method presented suggests that virtual prototyping can play a crucial role in making design decisions, even for designing medical devices with complex tissue–tool interfaces.

This section summarizes our previous study to provide a context to facilitate the reader in interpreting the design studies described in this paper. For a comprehensive description of the algorithm and implementation of dbd, the readers are directed to Ref. [1].

Design by Dragging is a software tool designed for exploring a large design space with multiple input parameters. A continuous design space is established by sampling solutions of complex finite element analysis (FEA) models. The algorithm of the dbd interface interpolates between these solutions when the user is interacting with the design space. The dbd features intuitive user interaction and data visualization, thereby enabling the user to approach design solutions in both forward (determining corresponding design outcomes of input parameter selections) and inverse (determining potential design parameter sets that will most closely yield the desired design outcomes) directions. Dragging is the main user activity required to explore a design space and manipulate design parameters. The internal algorithms interpret a user's intent based on his or her dragging operations. An interpretation can lead to various responses in the visualization such as modifying the geometry of the design or manipulating the output fields. Each drag operation returns new information about the design space, and the user is guided toward optimal solutions through step-by-step navigation. Such design exploration involves extensive interaction between the user and data. However, this interaction is conducted through minimal routine tasks. Human decisions are reflected in each stage of exploration to enable expert knowledge to inform the search process in each stage. Thus, the human designer is always well informed and in control of the optimal search process.

In the dbd environment, a user can perform drag operations directly on top of data visualizations. This type of interaction is called direct manipulation, and it enables both forward design and inverse design methods. Forward design allows a user to drag on a geometric feature to make design changes. The user can modify geometric parameters based on which geometric feature is being dragged and in which direction he or she wishes to explore additional solutions. The internal algorithm interprets the user's intent to decide which parameter to increase or decrease.

The focus of this study was to demonstrate the power of the inverse design of the dbd, which allows the user to directly manipulate a spatially distributed output field such as a stress field. The inverse design is enabled by the internal algorithm's continuous reshaping of the output field as the display morphs between the precomputed simulation results.

Figure 1 shows an example of traversing a design space for a breast biopsy needle system where a stress field is modified as a function of the input parameters. In the forward design, the user drags one edge of the opening window on the cannula and moves it toward the opposite edge to reduce the window length. In the inverse design, the user manipulates the stress field resulting from a perpendicular load applied to the needle tip. The goal is to determine the design alternatives currently having the highest stress region removed from the corner of the opening window to reduce the effect of a potential stress concentration failure.

A problem in designing a VAB tool was chosen as an example in this study. The design task was to find a balance between the tool's cutting performance, ergonomics, and component cost. To facilitate understanding of the trade-offs involved, the tissue cutting mechanism of VAB (Fig. 2) was investigated. The goal of this virtual prototyping was to determine the optimal solutions for motors, cutting speeds, and aperture size to assure proper tissue cutting in minimal time while maintaining an ergonomic handpiece weight and balance.

To initiate a cutting process, the inner cutter accelerates to the required rotational and translational speeds before coming into contact with the tissue. After contact, the inner cutter cuts through the tissue at a constant speed. In a successful scenario, the tissue would be continuously cut until a tissue sample was fully separated from the other breast tissue.

In this study, a two-step simulation process was developed to evaluate the performance of the VAB tool under various cutting conditions. In the first step, tissue cutting was modeled to predict tissue reaction force and torque, tissue deformation, and tissue fracture. The results were used in the second step to calculate the performance attributes of the tool for a complete operation cycle. For both steps, a simulation module was constructed to compute the required values. An FEA module was developed for the first step, and a matlab program was established for the second step based on motor equations. A flowchart illustrating the two-step process is provided in Fig. 3.

The following four design variables were considered as input parameters for the simulation: translational cutting speed (v), slice-push ratio (r), breast tissue type (b), and rotary motor selection (m). Notably, by controlling the first two variables, many cutting speed configurations for rotation and translation can be generated. The following five performance attributes were defined as output parameters for the simulation: total procedure time (t_p), tissue-sampling rate (s), maximal motor overload factor (K_max), mechanical system weight (w), and component cost (c). The t_p is related to the level of anxiety experienced by the patient during the procedure. The s is an indicator of cutting performance and the total volume of tissue extracted. The K_max dictates how much rest time between two sampling sequences is required for the selected motor to prevent overheating. The w is the total weight of all the design components in the handpiece and must be minimized for single-handed operation.

An explicit dynamic finite element model was developed to simulate the VAB procedure for extracting a breast tissue sample. At the initial time-step of the simulation, the needle aperture was filled with breast tissue, as the appropriate volume was drawn in by a vacuum. A finite element model of a local breast region, instead of the entire breast, was established. Figure 4 shows the device and tissue components in their initial assembly positions, including an outer cannula, an inner cutter, and a local breast tissue portion.

In the simulation, the inner cutter approached the tissue with preselected rotational and translational speeds. Following contact, the tissue gradually became deformed and fractured. The simulation input parameters were v, r, and b. Tissue deformation, tissue stress distribution, and tissue reaction torque T(t) were computed at each time increment. Among these output variables, T(t) became the input variable for the motor simulation module, whereas tissue deformation and tissue stress distribution were spatially distributed fields that were used to construct three-dimensional (3D) data visualization for direct manipulation.

This simulation computes the performance attributes of the VAB tool by analyzing a full operation cycle to retrieve five tissue samples. In each of the five sampling sequences, the inner cutter was under a varying load caused by T(t), which had to be overcome by the selected rotary motor as the driver of the inner cutter.

First, we considered each sampling sequence as a short-term motor operation. The K_max of the motor in a sampling sequence was calculated to determine a minimum resting period between two sampling sequences to prevent the selected motor from overheating. Subsequently, the overall tool performance attributes, t_p, s, w, and c were computed. A series of motor equations based on the motor selection guides of Maxon Motor [18] was implemented in the simulation module. The simulation input parameters were motor characteristics of m and T(t).

The levels of design variables were specified to generate 900 design points (Table 1). Three translational cutting speeds were selected, where the lowest being close to the clinical needle insertion rate and the highest being ten times faster than the lowest. The range of the r was specified to cover a maximum of rotational cutting speed of approximately 3000 RPM. In addition, the levels of r were nonuniformly distributed. A smaller increment of 0.1 was assigned when r < 2, because this range had been previously identified as a rapid change zone for the tissue reaction torque [12]. The increment was increased to 1 when 2 < r < 10. The aforementioned three breast tissue types were included. Five motor models were selected from the Maxon DC motor products with the following part numbers: 118,536, 315,174, 352,932, 226,748, and 110,066.

The 900 design points were used to sparsely populate a four-dimensional design space with over 10⁶ possible solutions (when an increment of 0.1 was used for v and r). The design points were simulated in a high-performance computing cluster provided by the Minnesota Supercomputing Institute. Each simulation job was run on an 8-core Sandy Bridge E5-2670 2.6 GHz processor. The 900 simulation output datasets were converted into NetCDF data format [19] to enable the dbd to efficiently compute design space sampling. Relative distances and smooth transitions (i.e., image warps) between all pairs of design points were calculated to provide interpolation between configurations and enable the forward and inverse designs.

The 900 design points were loaded into dbd to create a spatially populated design space. A warp was computed between each pair of design points to depict a smooth transition of the von Mises stress field from one point to another. Figure 5 shows the graphical user interface of dbd. A radar chart widget on the right-hand side displays the input and output design configurations. The spokes in the upper chart represent the input parameter selections, and those in the lower chart represent the output performance attributes. The minimal value of each parameter is at the center of the chart, whereas the maximal value is on the periphery of the chart. Each current design point and its corresponding design outcome set are indicated by red polygons. To explore the design space, a user can either drag a spoke of the input radar chart (top) to rapidly switch between design points or drag a spoke of the output radar chart (bottom) to specify a design outcome and search for matching design points. The larger rectangular area on the left is the visualization panel, which displays a spatially distributed output field with a regular view (bottom) and an enlarged sectional view (top). In the panel, the von Mises stress over the tissue region is displayed in both the regular and enlarged sectional views (shown at a cutting plane that separates both the outer cannula and the tissue into two symmetrical parts).

Before performing inverse design tasks, search criteria were set using the radar chart widget. All the parameters in the radar charts were assigned weights according to their relative importance to the overall design objective, thereby instructing the dbd system to compute relative distances between the design points that populate the design space in real time. Consequently, a parameter with a higher weight would be less likely to change in response to user interactions. Figure 6 shows the weighting condition used for all three study cases demonstrated later. In the input radar chart, tissue type and time-step were assigned an infinite weight to fix them at selected values (i.e., adipose tissue and final time-step), thereby enabling exploration of possible solutions while cutting adipose tissue and evaluating the resulting stress data after the cutting process. During the search, the other three design variables with a weight of zero were allowed to change freely. In the output radar chart, the three performance attributes with a weight of 1 were permitted only slight changes, whereas the other two with zero-weight tabs could change dramatically. Thus, the settings selected on the output radar chart widget ensured that design points returned in accordance with user drags on the stress visualization with c, w, and s values being similar to their currently set values. This illustrates a powerful decision strategy that is not available in traditional optimal design search processes where the user is not in the loop.

Studies have shown that slicing with pressing (linear displacement) is a more efficient way to cut soft materials than pure pressing [9,12]. In slicing with pressing, the cutting force is considerably reduced in the direction normal to the cutting surface, and material fracture is more likely to be caused by the shear force. This result is preferred for efficient VAB. The reduced cutting force in the axial direction causes the axial tissue displacement to decrease, enabling larger tissue samples to be obtained (as the FEA simulation results shown in Fig. 7). This is a favorable outcome for VAB tool design. Selecting an appropriate combination of translational and rotational cutting speeds is essential for producing sufficient shear force and reasonably low axial force. Furthermore, the choice of cutting speed also affects the requirements of the motor and weight and cost of the handpiece.

Inverse design was performed to reduce the axial displacement of the tissue for the design shown in Fig. 8(a), causing the tissue to be pushed too hard upon completion of the cutting process and forcing its leading edge to make contact with the base of the needle tip (the outer cannula and inner cutter were suppressed from view). With the weighting set in Fig. 6, we began to manipulate the stress visualization of the deformed tissue shape. We right-clicked on a location of the leading edge and three preview bubbles were displayed (an explanation of this feature is provided in Fig. 1). Notably, each preview bubble was an enlarged local region of a design point sought based on the predefined weighting. The middle preview bubble showed a design point with its leading edge shifted backward considerably. Subsequently, we viewed this design point by dragging to hit the center of the bubble (Fig. 8(b)). Finally, the new design point was reached, and the axial displacement of the tissue was shown to be largely reduced (Fig. 8(c)). Re-examination of the input radar chart suggested that the r should have increased from 0.7 to 4.0, whereas the v should have remained unchanged.

“Dry tap” refers to a failure of tissue sampling caused by a tissue sample not being fully separated from the other breast tissue. This problem was mentioned during our interviews with interventional radiologists. For the breast tissue to be severed by the cutter, either the maximum shear or maximum tensile damage criterion must be satisfied. We observed a “virtual dry tap” that occurred as a result of the tissue cutting simulations. In cases of virtual dry tap, no material damage criteria were met until the completion of cutting simulation. The tissue deformation finally became too large for the numerical solver to handle, and mesh penetration (i.e., the cutter penetrated the tissue rather than cutting it) occurred. Although the mesh penetration was a numerical error, it was also an indicator of dry tap, because it highlighted that the selected design was not capable of supplying sufficient force to reach the damage criteria required to break the tissue.

Dry tap was detected in a design point shown in Fig. 9(a). The strategy of solving this problem by using inverse design involved replacing the connected tissue (red) with empty space (blue). We right-clicked on a location in the red region, and five preview bubbles were shown near the click point (Fig. 9(b)). As previously mentioned, these preview bubbles were design alternatives sought by the dbd algorithm based on the predefined weighting. One preview bubble (circled in red) showed a possible design point for removing the tissue connection. We switched to the relevant design point for further examination, and the result showed that the dry tap problem had been resolved (Fig. 9(c)). Notably, this was not a unique solution. More design solutions could be found by requesting more preview bubbles from other right-click points. This case illustrates the power of this simulation tool to avoid selecting design parameters (motor selections in this case) that might result in the breast biopsy tool failing in the field of use.

The location of a concentrated stress region is usually of concern when evaluating FEA simulation results. For example, when designing a plate with a hole, a high-stress region can move around the hole depending on the loading condition. In this type of evaluation, the engineer must not only check the stress values but also examine the stress visualization to confirm the location and size of the stress concentration. The inverse design is a highly useful tool for such design activities. In this case, we demonstrated how to shrink a high-stress area through direct manipulation on visualization to find optimal design solutions. We hypothesized that cutting tissue with a more concentrated high-stress area surrounding the cutter–tissue interface enables a cleaner cut, which could help to produce high-quality tissue samples. Therefore, the size of the high-stress area surrounding the cutting tip was the performance indicator in this case. Figure 10(a) shows a starting design point with r = 0.4 and v = 50 mm/s. In this case, we evaluated the simulation result at an early cutting stage (the cutter had completed one-third of its total travel distance). The goal was to shrink and concentrate the high-stress areas (red regions). First, we right-clicked on a location in the lower red region to request preview bubbles for design suggestions. Two preview bubbles appeared near the click point (Fig. 10(b)). The left bubble indicated that the leading edge of the red region shifted to the lower left, which could lead to a narrower high-stress area. Thus, we switched to that design point and obtained a solution: r = 0.6 and v = 10 mm/s (Fig. 10(c)). This result suggests that by slightly increasing r, a cheaper design with reduced translational and rotational cutting speeds could yield better cutting performance. Such an insight cannot be obtained through design optimization tools that use currently available search methods, because the desired performance indicator is a visualized stress contour, which cannot be presented as a function explicitly or implicitly. However, by using the inverse design in dbd, we could directly manipulate the contour and rapidly reach a desired solution.

Following the previously reported development of the dbd software, this study applied the design methodology of the dbd to realistic optimal design problems of a VAB tool involving complex tool–tissue interactions. A total of 900 design points with various cutting conditions and motor choices were simulated. The simulation results were used to sparsely populate a large-scale, four-dimensional design space, and the dbd software was able to interpolate between these solutions in real time, guiding the user through the large design space.

We focused on the inverse design, which includes two essential tools: the radar chart widget and direct manipulation on visualization. The radar chart widget was used to assign weights to the input and output parameters, and the dbd system was instructed to compute the relationships between each pair of design points. By using the direct manipulation on visualization, we interacted with stress contours, used the drag function to move a high-stress region and leading edge of the tissue, and determined desired design solutions. In the first case, we improved cutting performance by dragging the leading edge of the tissue sample backward, thereby reducing the amount of total displacement of the tissue in the axial direction and providing a preferred solution where the tissue fracture was more heavily dominated by the shear force. In the second case, the dry tap problem was solved by dragging to grab the region with the connected tissue toward a design alternative. The design suggestion from this real-time exercise was to increase the slice-push ratio, while the translational cutting speed remained the same. In the final case, we determined a more effective design solution with a more concentrated high-stress region by shifting the leading edge of the region inward. A cheaper motor requirement for the handpiece was suggested to provide better cutting performance. All three cases illustrated processes involving humans in the design loop for direct interaction with the large design space, aided by the visualization. At each interaction step, the user's knowledge was integrated into the design process. In some cases, the design process adopted an active mode of driving the design away from a certain performance and toward the desired performance. In other cases, the user reacted to the visualization to observe something unexpected and used the preview bubbles to move to a different location in the design space. The insight gained through these interactions enabled the user to approach optimal and desired design solutions through the use of the powerful dbd software.

The authors would like to thank Michael Nelson, MD and his fellows who provided demonstrations of the VAB tool and shared their invaluable knowledge. The authors would also like to thank the Minnesota Supercomputing Institute for HPC resources and technical supports.

• National Science Foundation (NSF) (Grant No. IIS-1251069).

• National Institutes of Health (NIH) (Grant No. 1R01EB018205-01).

b =

breast tissue type

c =

component cost

K_max =

maximal motor overload factor

m =

rotary motor selection

r =

slice-push ratio

s =

tissue-sampling rate

t_p =

total time of a single biopsy sampling procedure

v =

translational cutting speed

w =

mechanical system weight