A Heterogeneous GPU–FPGA Computing Infrastructure for Efficient and Accurate Tumour Digital Twin Simulation
Precision medicine represents a paradigm shift in oncology, moving from population-level treatment protocols to patient-specific therapeutic strategies. Central to this vision is the ability to build and continuously refine tumour digital twins that integrate heterogeneous data—ranging from high-resolution pathology images to molecular and clinical parameters—and to use them to simulate drug response in silico. The CancerScan project is explicitly positioned within this transformation, aiming to enable scalable, reliable, and clinically actionable digital twins that can support oncologists in treatment selection and therapy optimisation.
However, the computational demands of this vision are unprecedented. Whole-slide pathology images reach gigapixel resolutions, parameter extraction pipelines are increasingly driven by deep learning models, and tumour simulations require the repeated execution of complex, multi-scale numerical models. Delivering actionable results within clinically meaningful time frames, while operating under strict energy and infrastructure constraints typical of hospital environments, is beyond the capabilities of conventional CPU-centric computing. In this context, heterogeneous hardware acceleration becomes a foundational enabler for precision medicine, rather than a performance optimisation.
Graphics Processing Units (GPUs) and Field-Programmable Gate Arrays (FPGAs) offer complementary capabilities that are essential to the CancerScan vision. GPUs provide massive parallelism and high throughput, making them ideal for computationally intensive tasks such as deep learning–based image analysis and large-scale parameter inference. FPGAs, by contrast, enable the implementation of highly specialised dataflow architectures with deterministic latency and exceptional energy efficiency. This makes them particularly suitable for embedding critical processing steps directly within medical devices or hospital-side infrastructure, close to data acquisition points such as pathology slide scanners. The synergy between GPUs and FPGAs enables a continuum of computation—from high-performance on-premise systems to low-power, edge-level processing—that aligns naturally with real clinical workflows.
Within CancerScan, the automated creation and execution of tumour digital twins require a tightly integrated computational framework that unifies image analysis, parameter extraction, and simulation. Despite significant progress in accelerator technologies, such an end-to-end framework does not currently exist. Existing tools remain siloed, often require specialised technical expertise, and are difficult to deploy at scale in clinical settings. A heterogeneous GPU–FPGA infrastructure offers the opportunity to overcome these limitations by enabling scalable, energy-efficient, and clinically deployable pipelines that can be seamlessly integrated into hospital environments.
To realise this vision, we will systematically characterise the computational workflow underlying tumour digital twin generation and simulation, identifying bottlenecks and constraints in terms of latency, memory footprint, and power consumption. Each computational component will be mapped to the hardware platform that best matches its requirements, partitioning the workload across GPU-accelerated high-performance systems and FPGA-based embedded solutions. This co-design approach ensures that the infrastructure evolves together with the medical models, preserving accuracy while meeting the stringent operational constraints of precision oncology.
Validation will be performed by comparing the outputs of the hardware-accelerated pipeline with those of a reference tumour simulator, which will serve as the gold standard, alongside system-level performance indicators, including execution time and energy efficiency. Success will be achieved when the proposed infrastructure demonstrates that patient-specific tumour digital twins can be generated and simulated with clinically acceptable accuracy, within practical time frames, and at sustainable energy costs.
By embedding heterogeneous hardware acceleration at the core of its computational strategy, CancerScan can move tumour digital twins from an experimental concept to a deployable clinical tool, paving the way for truly personalised, data-driven cancer treatment across Europe.
Author: Marco D. Santambrogio, Davide Conficconi
Links
Politecnico di Milano website: https://www.polimi.it/
NECSTLab website: https://necst.it/
NECSTLab LinkedIn: https://www.linkedin.com/company/necstlab/
Project website: https://www.cancerscanproject.eu/
LinkedIn channel: http://www.linkedin.com/company/cancerscan-project
X channel: https://x.com/CancerScan_eu
YouTube channel: https://www.youtube.com/@cancerscan-Project
Keywords
Tumour Digital Twins; Heterogeneous Computing; Hardware Acceleration; GPU–FPGA Co-Design; High-Performance Medical Computing; Energy-Efficient Architectures; Domain-Specific Accelerators: Scalable Healthcare Solutions; Sustainable Digital Health
