2.1.

Photodynamic Therapy

PDT, or the use of light activated photosensitizing drugs called photosensitizers (PS), is an approved therapy for various indications in oncology,¹¹ infectious diseases,¹² and other conditions.¹³ The interactions of light photons with the PS generates radicals, predominantly reactive oxygen species, resulting in spatially confined tissue destruction. The majority of the currently approved indications for PDT in oncology target organ surfaces, such as the skin,¹⁴ esophagus,¹⁵ or bladder,¹⁶ whereby access is provided by endoscopy and the tissue target depth is limited to 1 to 3 mm depending on the excitation wavelength. In these situations, typically individual, large surface covering light or photon emitters are used and the fluence rate gradient $[\mathrm{mW}{\mathrm{cm}}^{-2}{\mathrm{mm}}^{-1}]$ follows a one-dimensional distribution as a function of depth or radius from that emitter. Treatment of larger solid malignancies, such as in the brain, prostate, or pancreas, poses additional challenges related to adequate light delivery throughout the treatment volume. For interstitial PDT (iPDT), the clinical target volume often presents a complex three-dimensional (3D) shape and size, requiring placement of multiple sources with overlapping fluence rate distributions. To date, source placement and optical power allocation follow mostly empirical treatment plans such as the implementation of iPDT for malignancies in the central nervous system by Stummer et al.^17,18

2.2.

Prior Web- and Cloud-Based Monte Carlo Tools

Some web or cloud-based Monte Carlo (MC) simulators for biomedical applications have been published. CloudMc⁶ is a framework that allows an MC program to be launched on the Microsoft Azure cloud, with a portion of the photon packets executed on several different compute nodes, and final statistical results aggregated across nodes. This infrastructure is not applicable to our problem as it supports only Windows-based executables, and our simulation and optimization tools, like most open source projects, are instead developed on Linux. Wang et al.⁷ modify the ESG5 high energy electron and photon MC simulator to run a subset of the photons on each node in an AWS cluster; the lower-energy photons in PDT cannot be accurately simulated in this system. The most comparable work to ours is the MCX Cloud tool suite²⁴ which (such as our tools) simulates photon transport in biological tissues. Unlike MCX cloud, we use a public (AWS) cloud instead of a private cluster, and our tool suite includes not only Monte Carlo simulation of photon transport but also optimization of iPDT. Multiscattering^25,26 is a GPU-accelerated MC software with a web interface supporting quick and easy setup for users. It uses voxels to represent the problem geometry, rather than the tetrahedral mesh representation used by FullMonte and MMC, and has additional features for simulating spherical particles. Simulations are run on a cluster of three computers at Lund University, each of which has four GPUs.

3.1.

Photon Simulation: FullMonteSW

FullMonteSW⁵ performs Monte Carlo simulations of light propagation within 3D biomaterials of arbitrary shapes and tissue types to evaluate PDT treatment plans as well as for other light propagation studies in clinical¹⁶ and preclinical research.^27,28 The tool simulates the light propagation by tracing a user-specified number of photon packets²⁹ and aggregating statistics on photon absorption throughout the volume. It supports a variety of light sources, including point sources, radially emitting fibers, cut-end fibers emitting in a cone, arbitrary surface emitters, and combinations of these. Geometrically, it can model arbitrary 3D volumes with a tetrahedral mesh structure, allowing for accurate modeling of smooth and irregular surfaces, which are common in biological tissues. Finally, FullMonteSW is able to collect several user-defined output values, including the energy absorbed per tetrahedral element, and the energy entering or exiting internal or external surfaces. As shown in Fig. 1, FullMonteSW takes as input the optical properties of each tissue region, a 3D tetrahedral mesh that indicates each tissue region, the light source positions, their emission profiles, and their powers. The output data are used to either aid in iPDT optimization using the PDT-SPACE tool or directly visualize the fluence and the mesh structure using open-source tools such as Paraview.³⁰ Figure 2 shows an example of the visualization of an input and output mesh of an oral tumor at the base of a tongue with a single optical source.

Fig. 2

Example of the paraview visualization of an oral tumor at the base of the tongue with (a) a single point source inserted into the tongue and (b) the output fluence visualization from a light irradiation modeled for 633 nm.

FullMonteSW employs the Henyey–Greenstein scattering phase function, as the deep tissue tumor cases targeted employ lambertien or isotropic light sources. Since the code is open source, other scattering phase functions^31–33 can be added by end users if required for other applications. As described in Ref. 5, FullMonteSW has been cross-validated against two other tetrahedral Monte Carlo photon simulators (TimOS³⁴ and MMC³⁵) on 3D geometries, as well as MCML³⁶ on simpler layered geometries. FullMonteSW was also validated against analytic solutions using diffusion theory for two cases: an isotropic point source in a homogeneous infinite medium and a pencil beam in a homogeneous semi-infinite medium. FullMonteSW has internal consistency checks that are always active to ensure that physical laws such as conservation of energy are respected by the simulation outputs. FullMonte has a large regression suite comprising both system tests that check the final result accuracy and unit tests that validate internal software components. These tests are run and summarized automatically on every change to the software code; while the tests consume over 36 h of cpu time, they are run in parallel so they complete in $\sim 1.5\mathrm{h}$.

Apart from PDT, FullMonteSW can be used to simulate light propagation for other applications, such as disinfection of N95 filter facepiece respirators (FFR) using ultraviolet light germicidal irradiation,^37–39 and predicting the observed light distribution in imaging applications such as BLI imaging and trans-illumination spectroscopy for breast cancer detection.⁴⁰

3.2.

PDT Optimization: PDT-SPACE

The PDT-SPACE tool¹⁰ automates the treatment planning process by optimizing several degrees of freedom to achieve a desired tumor coverage while minimizing damage to the surrounding healthy organs at risk (OARs); its inputs and outputs are shown in Fig. 3. Given a set of light source locations and types, PDT-SPACE uses a convex optimization formulation to determine the power input to each source⁴³ to simultaneously maintain the tumor volume destroyed well above 98% while minimizing the OAR damage. In addition, PDT-SPACE can optimize the emission profile of cylindrical diffusers under the same goal while respecting user-specified manufacturability constraints.¹⁰ PDT-SPACE can also optimize the positions of the light sources to further improve the quality of the treatment plan. It treats the light source positions specified by the clinician as starting conditions and then uses a combination of simulated annealing and reinforcement learning to iteratively propose and evaluate improved probe positions.⁴⁴

Fig. 3

PDT-SPACE overall flow.

In each of these usage modes, PDT-SPACE must repeatedly call FullMonteSW to determine the light distribution throughout the tissue for the proposed plans. The program uses the result to evaluate plan quality against predetermined outcome conditions. Usually, ${10}^6$ photon packets are simulated in FullMonteSW to achieve sufficiently accurate results to guide optimization. PDT-SPACE also employs efficient search techniques to minimize the number of FullMonteSW calls and effectively the overall runtime.

3.3.

Open-Source Code and Docker Containers

FullMonteSW and PDT-SPACE are open-source and can be found on Gitlab repositories. Links to the repositories can be found on Ref. 45, and the wiki pages within the repositories provide extensive documentation on how to use and script the tools. These repositories also have extensive regression tests that are automatically launched whenever a source code change is committed to the repository, making it easier and safer for developers to update the code.

Both FullMonteSW and PDT-SPACE are complex programs that have dependencies on several software libraries and packages. Installing these dependencies can be daunting on some operating systems, so both tools support the alternative option of installing and running within a prebuilt docker container. Docker containers provide portable encapsulation of an environment to be run under any operating system, greatly reducing the time and effort required to get all the dependent software installed. The Wiki pages include instructions on docker installation and execution of the tools; the docker images are rebuilt and updated automatically whenever a source code change is committed so they are always up to date.

4.1.

Flow and Input Data

Figure 4 shows the usage and architecture of FullMonteWeb. User input is shown in orange, output is green, the open-source FullMonte tools are dark blue, whereas other open-source tools we leverage are shown in light blue. Users can upload a.vtk⁴⁷ format mesh that details the geometry of the tissues of interest or choose from a set of preloaded meshes. If an user is starting from medical imaging data, they can use tools such as ITK-snap⁴⁸ to segment the images then use the MeshTool available on the FullMonte website to create a tetrahedral mesh, which can be uploaded as input to FullMonteWeb. The FullMonteWeb web forms guide the user to enter the other inputs defining a problem (e.g., light sources, material optical properties, packet count to simulate, and desired outputs to compute). These forms validate that the entered data is self-consistent and physically possible and link to embedded tutorials that give examples of how to set up complete problems. Figure 5 shows a screenshot of one of the web forms used to set up a simulation; please see the Appendix for a more extensive walkthrough.

Fig. 4

FullMonteWeb overall flow.

Fig. 5

FullMonte simulation main page.

Entering the parameters defining a simulation in web forms is user-friendly, but it can limit the range of simulations possible unless a very large (possibly overwhelming) number of options are exposed in the forms. An alternative to option entry is scripting: the core solver, input and output routines in FullMonteSW can be directly called from scripts written in either the python or Tcl languages. This gives tremendous flexibility in problem set up by writing a new script that invokes the various input, output, and solver routines in different orders or even in loops. However, it can be a more difficult flow to learn as it requires some programming experience. To allow efficient use by both novice and power users, FullMonteWeb takes a hybrid approach. The most common options to FullMonteSW and PDT-SPACE are presented in web forms. Once all the forms have been completed, a Tcl script that implements the desired simulation or optimization flow is created and displayed as shown in Fig. 4. The script can be used as-is or the user can edit it in any way he or she desires. Editing the script allows specification of less common options, custom simulation flows, or loops that automate sweeping of parameters such as optical properties or other repetitive simulations required for statistical analysis. Once the Tcl script is finalized, it is uploaded to Amazon’s S3 storage and used to drive the simulation run on the AWS instance. Table 1 lists some of the options available directly in web forms, along with some of the power-user options and flows that Tcl scripts enable.

Table 1

Available options for simulation.

Fig. 6

AWS setup feature to allow scalable computing.

4.2.

AWS Integration

To leverage the power of cloud computing services, a user is required to launch an AWS Elastic Compute Cloud (EC2) instance (a virtual machine that allows access to some or all of the resources in a particular computer server) and provide its DNS name and.pem permission file to FullMonteWeb, embedded tutorials again guide the user through this operation. As shown in Fig. 6, FullMonteWeb automatically sets up this AWS instance. It performs a remote ssh (Secure Shell) operation to connect to the instance and scans the instance for the required environment and software (Docker, FullMonteSW, Paraview Visualizer, PDT-SPACE, as well as FullMonteCUDA⁴¹ and NVIDIA GPU drivers if the instance contains a GPU). FullMonteWeb will install only missing components on the user’s behalf, allowing simple operation but minimizing setup time if repeated simulations are run. Users can scale compute power not only by choosing a more powerful AWS instance but also by having multiple long-running simulations proceed in parallel by specifying and launching simulations on multiple AWS instances.

To ensure user data and compute resources are kept private, FullMonteWeb keeps a separate account for each user as shown in Fig. 7; sign-up is free. Recent simulation inputs and outputs are stored in the Amazon S3 storage cloud as shown in Fig. 4 and can be accessed through a history view in FullMonteWeb, as shown in Fig. 8.

Fig. 7

(a) Registration and (b) login pages to allow per-user compute and simulation data.

Fig. 8

History files download feature.

4.3.

Integrated Visualization

The outputs of a simulation are indicated in green in Fig. 4. They include report files detailing the dose-volume histogram (DVH) of a treatment plan for both the target tissue and the organs-at-risk; if plan optimization was performed with PDT-SPACE, the plan itself (probe locations, types and powers) is reported. Outputs such as the light fluence at each point in the 3D treatment volume are written into a vtk-format mesh file along with the mesh geometry itself so it can be visualized and processed with Paraview³⁰ or similar tools. FullMonteWeb also includes two integrated visualization tools so the key result data can be directly examined graphically in the web browser, without any installation of tools on the user’s computer.

4.3.2.

Interactive 3D visualization

The 3D visualization feature allows users to visualize the mesh and resulting fluence in 3D without the need for state-of-the-art graphics cards or specialized software on the user’s device. FullMonteWeb leverages the third party 3D visualization browser-based ParaView Visualizer software.⁴⁹ As shown in Fig. 4, the full Paraview application is automatically installed and run on the EC2 instance by FullMonteWeb. Paraview renders the 3D geometry and important outputs such as fluence in response to interactive user input, and the rendered data are sent to the Paraview Visualizer running on the user’s web browser for display. This arrangement keeps the more complex and computationally demanding software on the EC2 instance, with only the final display being performed in the user’s browser. The mesh and various tissue regions can be visualized before launching a simulation simply by clicking the “Open 3D Interactive Visualizer” button in FullMonteWeb; this can help in choosing the desired initial light source positions for a given case. After simulation completes, the resulting fluence can be visualized along with the tissue geometry. Figure 9 shows a few of the views that can be generated from a treatment plan output. The visualizer can also be used to review prior results stored in the S3 cloud or uploaded (in vtk format) by the user without rerunning the simulator.

Fig. 9

3D visualizer example views on a brain mesh and the resulting fluence.

4.4.

Efficiency versus EC2 Instance

One of the goals of FullMonteWeb is to abstract and configure compute resources so that researchers can choose the appropriate type and amount of CPUs, GPUs, and memory for any particular PDT simulation or optimization. There are usually two main metrics of concern to researchers: time to result and cost (while maintaining accuracy). We explore how the chosen EC2 instance affects these two metrics for simulations that vary in mesh size and whether a treatment plan is being evaluated with FullMonteSW or a plan is being created by the combination of PDT-SPACE and FullMonteSW.

Table 2 summarizes the six different EC2 instances we test. Each EC2 instance is a virtual machine with access to a certain number of virtual CPUs (vCPUs) and physical RAM. The Intel and AMD processors tested can run two threads on a single compute core (using a technology called hyperthreading or simultaneous multithreading) so Amazon counts a single core as two vCPUs. The instances we compare range in size from one vCPU to 64 vCPUs and have between 0.5 and 256 GB of physical RAM; two of the instances also include Nvidia GPUs. As Table 2 shows, the price per hour of an EC2 instance increases rapidly as its memory size and compute capabilities increase and the platforms we compare vary by over $500\times$ in cost per hour.

Table 2

EC2 instance details.

Table 3 details how the various EC2 instances perform on plan evaluation tasks using FullMonteSW simulations. For plan evaluation, we simulate three different meshes: a bladder mesh¹⁶ with a single point light source, the Colin27 brain atlas³⁵ (a low-resolution human brain mesh) with 1 point light source, and a high-resolution pig lung mesh—taken from Ramadan et al.²⁸—with light emitted from a complex surface source. Figure 10 shows a visualization of the three meshes. All simulations use ${10}^7$ photon packets, a typical value for accurate plan evaluation. The run times in Table 3 include all processing in FullMonteSW or FullMonteCUDA: loading of the mesh into memory, Monte Carlo simulation of ${10}^7$ photon packets, and generating output reports and an output vtk mesh for visualization. The time to set up the docker environment on the EC2 instance is 6 to 8 min but only occurs once and then is cached and reused, so it is not shown in Table 3. The time to upload an input mesh varies between 5 s (for the bladder case) and 1 min (for the pig lung case); once uploaded it is kept in persistent cloud storage so this is required only once per mesh, and hence it is not included in Table 3.

Table 3

Compute time and cost for a 107 photon packet simulation of a (i) bladder mesh, (ii) low-resolution human brain mesh, and (iii) pig lung mesh. See Fig. 10 for a visualization of the three meshes.

Fig. 10

The three models used in measuring the cost-runtime trade-off of the different EC2 instances. (a) A bladder mesh, (a) a low-solution human brain mesh (Colin27), and (c) a pig lung mesh.

The FullMonte runtime varies significantly across the meshes. The bladder case executes quickly because the photons have few interactions as they cross the transparent bladder void. The pig lung mesh executes slowly as photons frequently cross tetrahedral boundaries, requiring lookup of optical properties in the entered tetrahedron, and frequently cross tissue regions, requiring additional reflection and refraction calculations. Table 4 shows the relative performance of PDT plan optimization via PDT-SPACE using the Colin27 brain atlas and two different tumor cases, one small ($32{\mathrm{cm}}^3$ and three light sources) and one large ($104{\mathrm{cm}}^3$ and 19 light sources), which are cases T7 and T2 from Yassine et al.,⁵⁰ respectively. Note that PDT-SPACE currently supports only CPU-based implementations.

Table 4

Compute time and cost for PDT-SPACE optimization of two tumor cases, within the low-resolution human brain mesh. 107 photon packets are used in the required Monte Carlo simulations.

While the t2.nano EC2 instance (1 vCPU) is the slowest platform, it is also the most cost-effective platform for the FullMonteSW plan evaluations on the smallest (bladder) mesh. However, its low memory leads to poor performance on the (larger) low-resolution human brain mesh, and it cannot run the pig lung mesh at all, showing that larger instances are essential for some cases. GPU-enabled EC2 instances are the fastest platform for the two smaller meshes, with the fastest GPU (Nvidia Tesla V100) platform outperforming the fastest (64 vCPU) CPU platform by nearly $6\times$. For the bladder and the low-resolution brain mesh, the GPU-based instances are also more cost-effective than instances with a large number of vCPUs as their higher cost/hour is more than compensated for by their high performance, particularly for the less expensive Nvidia Tesla T4 platform.

On the other hand, the largest (pig lung) mesh, which contains 45.9 million tetrahedra, performs best on CPU-based platforms. This case has a high memory requirement of 28.1 GB and will not run on the two smallest CPU instances or the smaller GPU instance. The fastest platform is the 64 vCPU platform, and a high-end GPU platform of comparable cost is $1.3\times$ slower. This case has both a large mesh and a complex (surface) light source, resulting in a larger preprocessing time before the main Monte Carlo photon simulation begins. This time is reduced by a more capable CPU platform, as the preprocessing code does not leverage a GPU. While the GPU instance still outperforms even 64 vCPUs on the core Monte Carlo loop, it is not enough to overcome the slower preprocessing time due to its lower vCPU count. The most cost effective platform for this mesh is the 8 vCPU platform; its runtime is $2.1\times$ longer than the 64 vCPU instance, but due to its lower price per hour it is $4.8\times$ more cost-effective. This is again impacted by the higher preprocessing time in this test case. The mesh loading and preprocessing code is less parallel than the core Monte Carlo photon simulation loop, so the higher time spent in these operations in the pig lung case helps make the 8 vCPU instance the most cost-effective.

Table 4 shows that PDT-SPACE run time scales well as the number of vCPUs increases; on the large brain tumor case, a 64 vCPU instance reduces compute time by $10\times$ versus a 4 vCPU instance. The scaling is not perfectly linear partially because there is some serial code in PDT-SPACE and partially because the different vCPUs can compete for memory access in parallel code. While the largest (64 vCPU) instance is the fastest, the smaller (8 vCPU) is the most cost-effective, reducing the cost of a simulation by over $2\times$ versus the 64 vCPU instance.

The results show that both the fastest and the most cost-effective platform vary significantly with the problem being solved, highlighting the utility of users being able to choose from the over 100 types of EC2 instances to match compute to the problem at hand. Renting EC2 instances on demand also allows users to scale up their compute as needed, whereas buying and managing their own pools of servers is less flexible in adapting to demand.

5.1.

Optical Property Estimation

Using 3D CAD tissue phantoms, we leveraged FullMonteWeb to generate tissue optical properties look-up tables that correlate the anticipated light intensity at sensors or the extent of tissue necrosis for planned experiments for a range of diffuser depths inside the liver. Figure 11(a) shows a $6\times 12\times 6\mathrm{cm}$ cube, created using Autodesk Fusion 360 and converted into a.vtk format mesh, representing a liver tissue with a 2-cm long light diffuser placed at an oblique angle into the bile duct. The light diffuser is modeled as a separate region that emits photons diffusely (at a $100\mathrm{mW}{\mathrm{cm}}^{-1}$). Figure 11(b) indicates the fluence rate $[\mathrm{mW}{\mathrm{cm}}^{-2}]$ over the line shown in Fig. 11(a) for a range of ${\mu }_a$ and ${\mu }_{s^{\prime }}$ in liver tissue. A total of 11 ${\mu }_a[0.0005-0.0631{\mathrm{mm}}^{-1}]$ and 11 ${\mu }_{s^{\prime }}-[6-11{\mathrm{mm}}^{-1}]$ values were assigned to the liver tissue according to published values.^52,53 Thus, the simulation run comprises 121 optical properties executed in a single run. To perform this custom simulation, we leveraged the ability to edit the Tcl script generated by FullMonteWeb to add loops over optical characteristics; a simulation is performed for each possible pair of optical characteristics. FullMonteWeb executes the entire simulation, and the user can download a zip file with the output files for all 121 cases. Figure 11(c) shows the FullMonteWeb generated light detected at position $r=2\mathrm{cm}$ on the line shown in Fig. 11(a), as a function of ${\mu }_a$ and ${\mu }_{s^{\prime }}$. Utilizing multiple detector distances enables determination of the average tissue optical properties in the region of interest during subsequent in-vivo experiments in real time. It is important for the user to determine the required number of photon packets to be launched to achieve a sufficient fluence rate signal-to-noise ratio ($S/N$) for the highest ${\mu }_{\mathrm{eff}}$ and largest radial distance to be simulated.

Fig. 11

(a) Digital 3D liver phantom used for FullMonteWeb simulation, (b) exemplary dose attenuation curve along the solid white line shown on the phantom, and (c) simulation output showing the predicted fluence-rate at 2 mm from the center line as a function of the absorption and scattering coefficients.

Also, the disk storage size requested for these simulations given the number of tissue optical property combinations to be executed must be estimated as all simulation results are kept on disk on the EC2 instance prior to writing back to the Amazon S3 storage. Running a single test model indicated that the sufficient disk storage for one simulation is about 200 MB. Accordingly, a disk storage size of 64 GB was selected, to leave some room for required software installations. Simulated using ${10}^7$ photons on a m6i.16xlarge instance with 64 vCPUs, the overall run time of the 121 cases is 5 h and 11 min. The results can be visualized immediately on the EC2 instance and (if desired) individual result files can also be immediately downloaded to the end user machine. All the results are also compressed and transferred to persistent cloud storage (Amazon S3 storage) so they can be viewed from the FullMonteWeb history page (see Fig. 8) at any later date. For this large simulation, 15 GB of output data is produced, which requires 35 min to compress to a 5-GB.zip file, and a further 55 min to transfer to archival storage.

5.2.

Diffuser Placement Optimization

An example addressing refinement and reduction in preclinical subject usage pertains to optimizing the placement of optical diffusers within the vascular network in and around the pancreas, whereby the maximum permissible fluence needs to be determined to still permit individual scoring of the tissue response at each optical diffusers by pathology and histology. Here, FullMonteWeb was used to determine the location and power deliverable to all potential vessel locations and power per diffuser as part of a dose-escalation study in a porcine pancreas model. The fluence rates from multiple sites need to remain independent and the pancreas’ biological response at the various doses need to be quantifiable given the proximity of pancreatic tissue to a particular blood vessel, within a maximum amount of time permitted to occlude the vessel. An anatomically correct pig pancreas model was generated from contrast-enhanced CT images to execute dose-escalation simulations by constructing an in-silico 3D pancreas with the major blood vessels in and around the pancreas shown in Fig. 12(a). The model comprises three layers of the pancreatic tissue (shown in orange), the arterial blood vessel (in red), and the vein (in blue) with assigned optical properties based on the literature for aortic tissues.^53,54 The photon sources were modeled as cylinders with a 1 mm radius and 10 mm length. Figure 12(b) visualizes the resulting fluence around multiple sources placed simultaneously into the pancreas. The extent of the tissue destruction based on the photodynamic threshold model^55–57 as a function of fluence can thus be utilized for experimental planning. Simulations also provide the maximum power permissible to prevent thermal damage to the vessel wall, by limiting the fluence rate at the intima to $<300\mathrm{mW}{\mathrm{cm}}^{-2}$. Figure 12(c) shows a cross-sectional line between two illumination points. The fluence along this line shown in Fig. 12(d) illustrating that the maximum permissible intima surface irradiation is not breached and that while there is an overlap of the fluence rate profiles it occurs only after a fluence rate attenuation exceeding three orders of magnitude equivalent to more than three effective attenuation coefficients. It is up to the experimental planning to ensure that this fluence overlap will not result in an overlap in tissue necrosis, to enable evaluation of each illumination spot separately.

Fig. 12

(a) 3D model of pig pancreas (orange) with local arteries (red) and veins (blue), (b) pancreas cross section with color-coded fluence rate for two 10 mm long diffusers embedded in the superior mesenteric artery, (c) visualization of fluence perpendicular to the light emitter, and (d) the fluence rate intensity distribution along the center after establishing the maximum permitted power density to the intima.

7.1.

Treatment Plan Evaluation

To run FullMonteSW on FullMonteWeb, a series of input pages guide users intuitively to provide the necessary files and parameters for simulations. As shown previously in Fig. 5, the first page prompts the basic information, including the input mesh file, mesh units, types of energies to record, and more. In the back-end, FullMonteWeb takes all the selections and generates a Tcl script that runs FullMonteSW on the provided EC2 virtual machine.

The second page prompts for the mesh’s region properties and allows users to save certain regions as presets for future quick selections. As shown in Fig. 13, advanced users can download the semifinished Tcl script from this page, manually complete the script, upload it back to FullMonteWeb, and skip the rest of the setup procedures.

Fig. 13

Advanced options for FullMonte simulation.

The third page asks for the light source types and positions. The 3D visualization feature providing a visual model of the input mesh and its surrounding space aids users in selecting light source locations. The confirmation page provides a chance for users to review the settings and make any final changes to the Tcl script before launching the simulation. All Tcl scripts downloaded from these steps are prepared with comments to guide the user toward completing it.

Upon confirmation, a progress page appears providing feedback on the execution progress in real time, as shown in Fig. 14. Upon completion, the simulation logs page is presented, providing potential warnings and errors that occurred during the simulation. Users can visualize the results, or download the Tcl script, input files, and output files via the simulation history page.

Fig. 14

FullMonte simulation progress.

7.2.

Treatment Plan Creation

To run PDT-SPACE, the user is guided by multiple input pages to provide the necessary files and parameters for software execution. A Mosek⁵⁸ license needs to be obtained and uploaded to PDT-SPACE to access its Fusion C++ API library for solving the convex optimization problem. The license can be obtained by following the link provided on the upload page.

Next the user is asked to specify the input mesh, optical and issue property files, the number of photon packets, wavelength, and the tumor weight. The tumor weight is required when optimization is to be executed to maximize tumor destruction while minimizing OAR damage. Experienced users can specify the pruning normalization factor to scale the light dose thresholds. Subsequently, the user can select the source type, placement type, and upload an initial placement file containing this information. In the back-end, FullMonteWeb takes all the settings and generates a parameter script that contains all input parameter names and values for running PDT-SPACE on the specified EC2 instance.

To launch PDT-SPACE optimization, FullMonteWeb executes two scripts that are created during AWS setup. The first script is to download and open the docker image, which contains the PDT-SPACE library on the user’s AWS account. The second script is to set the running environment and execute the PDT-SPACE command with the auto-generated parameter file. Once confirmed, users will be redirected to a progress page, as shown in Fig. 15.

Fig. 15

Progress page when running PDT-SPACE.

Once the optimization completes, the website will automatically redirect users to a results page, as shown in Fig. 16. The page contains total energy, number of packets, fluence distribution across all material regions, relative power allocation for the specified light sources, and optimization run time. Users can then visualize the PDT-SPACE output or download the log files from the link at the bottom of the page.

Fig. 16

An example results page on FullMonteWeb after running PDT-SPACE.

7.3.

Robustness and Security

To ensure that users can only access their own files, all user data are protected by their accounts. As shown in Fig. 17, if users attempt to run FullMonteSW, PDT-SPACE, visualizers, or to access the simulation history page without logging in, they will not be able to proceed.

Fig. 17

Error when user attempts to use core functions without logging in.

Since FullMonteSW, PDT-SPACE, and visualizer computations cannot be done in the browser, users are required to set up an AWS virtual machine. If FullMonteWeb cannot detect a valid virtual machine setup on the account, the user will be redirected to the AWS Setup page, as shown in Fig. 18.

Fig. 18

Error when user attempts to run simulation, optimization, or visualization without setting up a valid EC2 virtual machine.

To ensure the website’s reliability against invalid inputs and mismatches between mesh and specified material properties, there are several restrictions and validation modules implemented in the back-end. For example, the refractive index should be at least 1, so if the user specifies a value outside of this range, they will be alerted, as shown in Fig. 19. Furthermore, mesh files are typically specified in units of centimeters or millimeters, so users are required to specify the mesh unit when they upload the mesh file. All inputs following the upload should be in the same unit, so the website automatically converts all preset information to the same unit.

Fig. 19

Error when user enters an invalid input.

7.4.

Visualization

7.4.1.

Dose-volume histogram

When the visualization feature is launched after a successful FullMonte simulation, the user will be prompted to upload a tissue properties file containing the threshold fluence for each region. FullMonteWeb will then automatically generate a DVH and display it on the web page. Likewise, a DVH will also be generated after a successful PDT-SPACE run but without the step of uploading a tissue properties file (see Fig. 20). A DVH serves to show the distribution of fluence (energy intensity) throughout each body tissue region. With this graph, the user can understand the volume of healthy tissue and tumor after receiving a certain percentage of its dose threshold. From the figure, the $x$ axis represents the percentage of fluence dose with respect to the dose threshold for each tissue in the mesh and the $y$ axis represents the percentage of a region that receives at least a given dose.

Fig. 20

A dose volume histogram generated from PDT-SPACE output of an iPDT plan for a brain tumor with a $32{\mathrm{cm}}^3$ volume in the frontal lobe.