Monte Carlo Shows the complexitites and 10 year evolution
Lilhge
2009
https://www.eecg.toronto.edu/~jayar/pubs/lo/lojbo09.pdf
Conclusion
Using the MCML program as the gold standard, custom pipelined hardware designed on a multi-FPGA platform known as the TM-4 achieved an 80 times speedup compared to a 3-GHz Intel Xeon processor. The development time was approximately 1 person-year and future modifications can be readily implemented due to the use of a modularized pipelined architecture. Isofluence distribution maps generated by FBM and MCML were compared at 100 million photon packets, showing only a 0.1 mm shift in the hardware-generated isofluence lines from those produced by MCML for fluence levels as low as 0.00001 cm−2. This shift is negligible within the context of PDT treatment planning considering the typically much larger margin of safety for surgical resection or treatment planning in radiation therapy. 8 Implications and Future Work The limitations of the current prototype design, such as the number of layers, could be relaxed on newer FPGA platforms, which offer more on-chip memory and other resources. Migrating the current design to modern Stratix III FPGA chips will result in a projected 240 times speedup, requiring minor modifications to the communication interface. In future studies, the use of external memory will have several implications. First, more replicas of the design can be accommodated since the on-chip memory space is a limiting factor, directly translating to an increase in the attainable speedup. Second, using external memory enables the 3-D modeling of tumors, which for realistic cases would require at least 10241024 1024 voxels a minimum of 4 GBytes assuming 4 bytesper voxel. Finally, the significantly larger memory space offered by external memory will enable further optimization of the number of entries in the look-up tables to improve the accuracy of the simulation. Determining the precise trade-offs
between accuracy and resource usage as well as the migration to newer platforms will be the subject of future work. For investigators interested in accelerating other light
propagation models such as FEM-based models that solve the radiative transfer equation numerically using the diffusion approximation,29 an FPGA-based approach may serve as an
alternative. Here, the unique technical challenges will primarily include mapping the matrix operations onto hardware and implementing an iterative solver based on techniques such as
the conjugate gradient method.30 Tailoring the FPGA-based hardware to the system of matrices specific to the application will be a key step in the design process.
The possible implications of this study are twofold. First, the pipelined design could form the basis on which more complex MC simulations or other light transport models can be built. The flexible pipelined architecture enables the addition of extra stages such as those required by external memoryaccesses without significantly impacting the performance.
Secondly, the dramatic reduction in treatment planning time achieved by an FPGA platform may potentially enable realtime treatment planning based on the most recent images of
the treatment volume, taking into account the changing tissue optical properties as the treatment progresses. Currently, pretreatment models assume constant values for tissue optical properties and ignore the dynamic nature of tissues, which could directly affect treatment outcomes in interstitial PDT.31 The significant performance gain provided by the hardware approach can potentially enable PDT treatment planning in heterogeneous, spatially complex tissues using more sophisticated MC-based models.
2015
https://www.researchgate.net/publication/276442712_Treatment_plan_evaluation_for_interstitial_photodynamic_therapy_in_a_mouse_model_by_Monte_Carlo_simulation_with_FullMonte
CONCLUSIONS AND FUTURE WORK5.1. CONCLUSIONSWe have shown that Monte Carlo methods are computationallytractable for PDT treatment evaluation, permit continuously-variable quality-runtime tradeoffs, and are amenable to hardwareacceleration. Additionally, the computation of dose-volume his-tograms (DVHs) from MC simulations functions as a signicantvariance-reduction scheme which greatly reduces the runtimerequired to achieve a given level of output condence. Currentrun times can be sufciently low, even on a laptop computer,that tetrahedral-mesh MC is a viable option for evaluating PDTtreatment plans. Higher-precision simulations with a larger num-ber of packets which will permit accurate sensitivities (e.g., toprobe placements or to personalized optical properties) can alsobe calculated. For iterative treatment planning and algorithmicoptimization, accelerated hardware enables a very large volumeof simulations to be run quickly. In short, given an appropriateunderstanding of the requirements for acceptable output quality,and of the effects of DVH generation, Monte Carlo simulationsfor PDT treatment planning may be considerably (integer factors)lessexpensivethanisgenerallyperceived.5.2. OPTICAL PROPERTY VARIABILITYThis work demonstrates the capability and low computationalcost of using Monte Carlo methods to evaluate the dose deliv-ered by a given light source conguration into a specied tis-sue geometry with known optical properties. In the practiceof interstitial PDT delivery, however, patient optical propertiesvary widely [33, 34]. Development of treatment plans whichare robust to a range of optical properties, as well as tech-niques for in-vivo detection and compensation for such variability(buildingonprioreffortsintheprostate[4, 25]), would beimportant contributions to PDT practice. The rst part involv-ing assessment of variability is enabled in a straightforward wayby the present work, requiring only simulation of clinically-derived meshes over a range of properties with analysis of theresults. Development of robust optimization techniques andoptical property detection/compensation methods are more dif-cult questions; this work provides an important prerequisite,namely a fast and accurate plan evaluation engine.5.3. ALGORITHMIC TREATMENT PLANNING FLOWUsing the infrastructure developed, we aim to create anautomated numerical optimization system which works towarduser-dened treatment plan goals (minimum dose for tumor,maximum dose to organs at risk, and a penalty function for devi-ations). It will need to be robust to misdirection due to the out-put variance, though to what degree remains to be determined.One may contemplate a highly robust algorithm (e.g., simulatedannealing) which works on very short, low-quality/high-variancesimulation runs; or, one may choose to employ an algorithmwhich is less demanding in the number of sample plans toevaluate but requires higher-quality input.5.4. HARDWARE ACCELERATIONIn previous work, we have also demonstrated the capabil-ity to perform identical Monte Carlo simulations using Field-Programmable Gate Arrays (FPGA) custom digital logic [23],running 4x faster and 67x more power-efciently than a high-end quad-core Intel processor for mesh sizes up to 48k elements.Such simulations can replace the software kernel used herein togenerate results considerably more quickly and cheaply than thegeneral-purpose CPU used. We aim to use this specialized plat-form to enable algorithmic optimization of PDT treatment plans.5.5. DOSE CONCEPT REFINEMENTWhile the present work focuses entirely on delivering a evaluat-ing the light uence delivered, denitions of photodynamic doseencompassing more factors exist [5, 6]. Given the existing sim-ulator’s output of uence within the tissue, additional explicitdosimetric information regarding oxygenation, photosensitizerconcentration, and tissue sensitivity can be incorporated to pro-duce a DVH of photodynamic dose which should more closelypredict the clinical outcome. Such additional factors would needto be modeled separately and integrated with the uence informa-tion from the optical simulator, for computation of photodynamicdose-volume histogram rather than (or in addition to) simplelight uence dose.5.6. IMPACT OF MESH REFINEMENTThe connection between mesh size and DVH variance needs fur-ther investigation. As discussed in the results section, there arecontrary pressures when mesh element size decreases: individ-ual element variance would tend to increase because of a smallernumber of arrivals to be summed; DVH variance would how-ever tend downwards due to the variance-reduction properties ofsorting the uence elements. Mesh renement also adds a mildcomputational burden, which may or may not be offset by thedecreased number of packets which will be required to achieve anacceptable level of condence in the DVH
Recent:
https://www.soscip.org/2017impactreport/impact-stories/fast-and-accurate-biophotonic-simulations-for-personalized-photodynamic-can/
https://www.oce-ontario.org/docs/default-source/Presentations/6-vaughn---soscip_oce_forum.pdf?sfvrsn=2&sfvrsn=2
https://onlinelibrary.wiley.com/doi/full/10.1002/jbio.201800153
https://slideplayer.com/slide/12717298/
Make PDT simulation both fast and accurate 2. Leverage fast simulator to enable automatic optimization of plan, and robustness evaluation with tissue variation 3. Develop full flow from MRI data through simulation to visualization 4. Increase PDT efficacy, and broaden applicability to more cancer indications
Research team: – Betz: compute acceleration, software development – Lilge: medical physics modeling, PDT expertise – Together: develop complete software + hardware flow IBM – Agile hardware, APIs, and simulation environments – Technical support and expertise on agile platform – Interaction with IBM researchers working on related medical simulation and imaging problems Theralase – PDT and optical measurement expertise – Calibrate simulator against measured light / PDT results
Theralase
--Provider of PDT photosensitizers and lasers: better PDT treatment planning larger market
--Evaluate a treatment plan in silico: can tell if PDT plan will lead to a good result before treatment
--Automatic optimization: create a better and more robust treatment plan, quickly
PDT and optical measurement expertise – Calibrate simulator against measured light / PDT results
Maybe that's where the market projections are coming from?
Are we about to displace DUSA - ALA, Photofrin and others??
Achievements to Date
World’s fastest software simulator for general Monte Carlo simulation of light
Agile hardware implementation in progress
– Single compute pipeline functions in simulation
• But still debugging in actual hardware
– 4x performance of high-end, four-core CPU
– Expect ~16x performance and 60x power-efficiency for scaled-up
full system
Prototype components of most of MRI image
simulation visualization flow