Skip to content
Murali RaviMedical Instrumentation
Deployed2016-2019

Hardware Acceleration for Medical Imaging

Making the best reconstruction algorithm fast enough to actually use.

Iterative algorithms like Maximum Likelihood Expectation Maximization produce the highest-quality tomographic images and take the longest to compute: which is precisely why clinics often cannot use them. This work treats that as an engineering constraint rather than a fact of life, implementing a dedicated acceleration framework across FPGA fabric and DSP vision pipelines.

0×

MLEM speed-up on FPGA over optimised Xeon software

FPGA acceleration (MLEM)
288×
DSP processing acceleration
250%
Xilinx VC709 platform
Virtex-7
Citations, IEEE Access 2019
44
MAXIMUM LIKELIHOOD EXPECTATION MAXIMIZATION: ITERATIVE LOOPForwardprojection1Compareto measured2Backprojection3Updateestimate4× N iterationsUNTIL CONVERGENCETHE UPSIDEHighest imagequality availableNoise modelling, better contrastTHE COSTExecution timeblocks clinical useEvery iteration re-projectsThe trade-off is not a law of nature: it is an architecture problem.

Summary

What the work set out to do, and the reasoning that shaped it.

This research addresses the computational bottleneck in tomographic medical imaging. While iterative algorithms like Maximum Likelihood Expectation Maximization (MLEM) provide high-quality images, their long execution times often prevent real-time clinical use.

The response was a dedicated hardware acceleration framework using FPGAs and DSPs: moving the mathematics off general-purpose processors and onto fabric designed to exploit its parallelism.

The FPGA work implemented a parallel structure for MLEM on a Xilinx Virtex-7 VC709 using High-Level Synthesis, achieving a 288× speed-up compared to optimised software execution on an Intel Xeon workstation. Separately, a dual-core Blackfin ADSP-BF609 with an embedded Pipelined Vision Processor delivered a 250% acceleration for clinical post-processing tasks.

Impact

What changed as a result.

01

A 288× speed-up over optimised Xeon software execution moves MLEM reconstruction from a batch operation into the interactive regime.

02

The DSP framework accelerated diagnostic calculations: Thyroid Uptake Ratio and Glomerular Filtration Rate among them: by 250%, addressing the post-scan analysis path rather than only reconstruction.

03

Together, the two frameworks facilitate real-time diagnostic imaging, allowing rapid access to medical images during critical, time-sensitive healthcare procedures.

04

The IEEE Access paper documenting the FPGA architecture is the most-cited output of the research programme.

Figures

The work, drawn.

Purpose-built diagrams: each one carries an idea that prose alone would take a page to deliver.

MAXIMUM LIKELIHOOD EXPECTATION MAXIMIZATION: ITERATIVE LOOPForwardprojection1Compareto measured2Backprojection3Updateestimate4× N iterationsUNTIL CONVERGENCETHE UPSIDEHighest imagequality availableNoise modelling, better contrastTHE COSTExecution timeblocks clinical useEvery iteration re-projectsThe trade-off is not a law of nature: it is an architecture problem.
01

The MLEM bottleneck

Forward projection, comparison, back projection, update: repeated until convergence. Quality comes from iteration; so does the delay.

SEQUENTIAL: INTEL XEON WORKSTATIONvoxel block 1voxel block 2voxel block 3voxel block 4ONE AFTER ANOTHER · TIME ACCUMULATESPARALLEL: XILINX VIRTEX-7 VC709PROCESSING ELEMENTS EXECUTE CONCURRENTLYDESIGN FLOWAlgorithmMLEM in C/C++RestructureExpose data parallelismHLSHigh-Level SynthesisRTLGenerated fabricVirtex-7VC709 deploymentKEY INSIGHTMLEM's forward and back projection steps hide substantial data parallelism: exposing it is what makes the fabric worth using.
02

Parallel architecture on Virtex-7

Restructuring MLEM to expose its latent data parallelism, then mapping it onto FPGA fabric via High-Level Synthesis.

RELATIVE THROUGHPUT: LOGARITHMIC SCALE10×100×288×Intel Xeon workstationOptimised software execution1× baselineADSP-BF609 + PVPClinical post-processing250% fasterXilinx Virtex-7 VC709Parallel MLEM via HLS288×FPGA: RECONSTRUCTION288×SPEED-UP VS OPTIMISEDXEON SOFTWARE EXECUTIONDSP: POST-PROCESSING250%ACCELERATION ON THYROIDUPTAKE RATIO & GFR
03

288× against optimised software

Measured against optimised software execution on an Intel Xeon workstation: not against a naive baseline.

ANALOG DEVICES ADSP-BF609: DUAL-CORE BLACKFIN WITH PIPELINED VISION PROCESSORADSP-BF609Core 0BlackfinCore 1BlackfinPipelined Vision ProcessorHARDWARE IMAGE PIPELINEL1 / L2MEMORYACCELERATED CLINICAL CALCULATIONSThyroid Uptake RatioActivity quantification from acquired countsGlomerular Filtration RateRenal function derived from time-activity dataRESULT250% accelerationRECONSTRUCTION (FPGA) → POST-PROCESSING (DSP) → CLINICAL READOUTTwo accelerators, two problems: the FPGA rebuilds the image, the DSP reads it.
04

Pipelined Vision Processor path

The ADSP-BF609's PVP handles clinical post-processing: uptake ratios and filtration rates: at 250% of software speed.

Methods

How it was done.

Parallel MLEM on FPGA fabric

MLEM's iterative structure hides substantial data parallelism inside its forward and back projection steps. Restructuring the algorithm to expose that parallelism is what makes an FPGA implementation worthwhile rather than merely different.

High-Level Synthesis

HLS was used to target the Xilinx Virtex-7 VC709, allowing the parallel architecture to be expressed and iterated at a higher level of abstraction than register-transfer design.

Pipelined Vision Processor

The Analog Devices ADSP-BF609's embedded PVP provides a hardware pipeline for image operations, used here for clinical post-processing rather than reconstruction.

Diagnostic calculation offload

Thyroid Uptake Ratio and Glomerular Filtration Rate calculations were moved onto the accelerated DSP path, producing a 250% improvement in processing speed.

Specifications

The numbers.

FPGA platform
Xilinx Virtex-7 VC709
Design methodology
High-Level Synthesis (HLS)
Algorithm
Maximum Likelihood Expectation Maximization (MLEM)
FPGA speed-up
288×vs. optimised software on Intel Xeon workstation
DSP platform
Analog Devices ADSP-BF609Dual-core Blackfin
DSP co-processor
Pipelined Vision Processor (PVP)
DSP acceleration
250%
Clinical calculations
Thyroid Uptake Ratio, Glomerular Filtration Rate

Development timeline

From simulation to the clinic.

2016

System-on-Chip formulation

A novel System on Chip design for thyroid imaging studies presented at IEEE ICAECCT: the first formulation of the accelerated pipeline.

2017

Low-cost real-time platform

A practical design and implementation of a low-cost platform for real-time diagnostic imaging published in IEEE Access.

2018

DSP acceleration framework

ADSP-BF609 with Pipelined Vision Processor applied to clinical post-processing, achieving 250% acceleration on diagnostic calculations.

2019

FPGA parallel architecture

Parallel MLEM on Xilinx Virtex-7 VC709 via HLS achieves 288× speed-up. Published in IEEE Access; the programme's most-cited work.

Outcomes

  • 288× MLEM reconstruction speed-up on Xilinx Virtex-7 VC709 FPGA
  • 250% acceleration of clinical post-processing on ADSP-BF609 with PVP
  • Two IEEE Access journal papers (2017, 2019)
  • IEEE ICAECCT conference paper on System-on-Chip thyroid imaging (2016)
  • Real-time diagnostic imaging made feasible for time-sensitive procedures

Collaborators & support

  • Sri Sathya Sai Institute of Higher Learning
  • A. Sewa
  • S. S. S. Sanagapati
  • T. G. Shashidhara
  • S. Sivaramakrishnan

Funded under

DST · INR 96 Lakhs

2016-2021 · Project Associate