The Educational Challenge
Newly qualified radiographers and trainee radiologists face a steep learning curve when they begin clinical MRI practice. The relationships between TR, TE, TI, flip angle, bandwidth, matrix size, field of view, and the resulting image contrast are complex, interdependent, and often counter-intuitive. Traditional teaching methods — lectures, textbooks, observation — can describe these relationships, but they cannot replicate the experience of sitting at a console and watching an image change in real time as each parameter is adjusted.
Scanner time is expensive, clinically allocated, and rarely available for teaching purposes. Even when training slots can be arranged, the learning opportunity is limited by the need to work with real patients and the pressure to produce clinically acceptable images. There is no opportunity to set deliberately poor parameters, observe the result, and understand why it failed.
Our simulator removes these barriers entirely.

Capabilities

Pulse Sequence Library

The simulator provides six complete pulse sequence implementations, each with correct signal equations and clinically appropriate parameter ranges:

• Spin Echo (SE) — the foundational sequence for understanding T1, T2, and proton density weighting through TR and TE manipulation.
• Turbo Spin Echo (TSE) — with echo train length control, phase-direction blurring simulation, and magnetisation transfer contrast effects that realistically reduce grey–white matter differentiation at high ETL.
• Gradient Echo (GRE) — implemented as three distinct sub-types: Spoiled (SPGR/FLASH/T1-FFE), Rewound (GRASS/FISP/FFE), and Balanced SSFP (TrueFISP/FIESTA/b-FFE), each with its own steady-state signal equation and characteristic contrast behaviour.
• Inversion Recovery (IR) — with adjustable TI for selective tissue nulling. Presets automatically calculate field-strength-appropriate null points for white matter and grey matter.
• STIR — with field-strength-adaptive fat nulling and educational warnings about contraindication with gadolinium contrast.
• FLAIR — with field-strength-adaptive CSF nulling for periventricular lesion conspicuity.

Comprehensive Parameter Control

Every clinically relevant acquisition parameter is independently adjustable, giving students full control over the signal-to-noise, spatial resolution, and scan time trade-offs that define real-world protocol optimisation:

• Timing parameters: TR (2–10,000 ms), TE (1–300 ms), TI (10–5,000 ms), flip angle (1–90°), echo train length (2–32).
• Spatial parameters: FOV (100–500 mm), rectangular FOV (50–100%), phase encoding direction, frequency matrix (64–512), phase matrix (64–512), slice thickness (1–10 mm).
• Signal quality: NEX (1–4), parallel imaging acceleration (R = 1–4) with g-factor noise penalty, receiver bandwidth (50–500 Hz/pixel).
• Clinical options: frequency-selective fat saturation, intravenous gadolinium contrast simulation with flow void modelling, no phase wrap oversampling.

Multi-Anatomy, Multi-Field-Strength Design

The simulator includes accurately segmented tissue maps for two anatomical regions — axial brain (eight tissue types including cortical bone, CSF, grey matter, white matter, fat, muscle, skin, and blood) and sagittal knee (nine tissue types including articular cartilage, synovial fluid, ligament, cortical bone, and trabecular bone marrow). Each anatomy can be imaged at 0.5T, 1.5T, or 3.0T, with a complete tissue relaxation database providing field-strength-dependent T1, T2, and proton density values derived from the published topic literature.
More anatomical areas are scheduled to be added shortly - and all upgrades will be accessible to our participants over the duration of their access period.

Artefact Simulation

Understanding artefacts is essential for clinical practice, yet difficult to teach without hands-on experience. Our simulator can reproduce all of the the major acquisition-related artefacts encountered in routine imaging including:

• Phase wrap-around — visible when the phase FOV is smaller than the anatomy, and removable with the no-phase-wrap oversampling toggle.
• Chemical shift — bandwidth-dependent spatial misregistration at fat–water interfaces, correctly scaled to field strength and displayed in the frequency-encoding direction.
• TSE blurring — T2-dependent signal decay across the echo train causing high-spatial-frequency attenuation in the phase direction.
• Realistic noise (and noise spike artefacts) — signal-dependent noise modelling that responds correctly to voxel volume, NEX, bandwidth, field strength, and parallel imaging acceleration.

Our course covers 26 different artefacts - it's better that our new course participants can learn about these disasters on the simulator rather than on real patients during a busy list!

Protocol Optimisation Engine

The Optimise engine embedded within the simulator encapsulates the protocol design expertise that typically takes years of clinical experience to develop. For any given combination of pulse sequence, image weighting, anatomy, and field strength, the engine calculates a complete optimised protocol that balances diagnostic image quality against acquisition time — the fundamental trade-off at the heart of every MRI examination.
The optimisation considers the full parameter space: timing parameters for contrast, spatial parameters for resolution and coverage, receiver bandwidth for SNR and chemical shift control, parallel imaging for scan time reduction with appropriate g-factor trade-off, signal averaging for thin-slice SNR compensation, phase encoding direction for anatomy-appropriate wrap avoidance, and fat saturation where clinically indicated.
When the student engages with the built-in Learning Mode function, they can build a sequence/protocol from scratch - test the image appearance with every change and when they are done, the feedback report provides specific, educational commentary on every parameter that differs from the optimised protocol. This commentary explains not just what should change, but why — linking each adjustment to its physical consequence in terms of contrast, resolution, SNR, artefact behaviour, or scan time. Estimated scan times are displayed for both protocols, quantifying the time penalty of suboptimal parameter choices.

Technical Specification

Our simulator was designed an built in-house by Dr John Talbot, whose doctoral thesis was in the field of technology enhanced learning. It is delivered as a browser-based application requiring no software installation, plug-ins, or dedicated hardware. It runs in any modern web browser on desktop, laptop, or tablet devices. Image rendering is performed client-side in real time, providing instantaneous visual feedback as parameters are adjusted. The tissue maps, relaxation database, and edge detail overlays are served from a secure, authenticated server environment integrated with the MRI in Practice online learning platform.

Physics fans only…

SIGNAL MODEL

The Talbot 3000 computes signal intensity per voxel from the Bloch equations using tissue-specific T₁, T₂, and ρ values drawn from a field-strength-indexed relaxation database (0.5 T, 1.5 T, 3.0 T) compiled from Stanisz et al. (2005), de Bazelaire et al. (2004), and Gold et al. (2004). Six pulse sequences are implemented with closed-form steady-state signal equations: SE (ρ(1−e⁻ᵀᴿ/ᵀ¹)e⁻ᵀᴱ/ᵀ²), TSE with effective-TE echo-train modulation and MT attenuation coefficients, three GRE sub-types (spoiled: Ernst-angle steady-state; rewound: √(T₂/T₁)-dependent SSFP−FID; balanced: T₂/T₁-weighted bSSFP with near-unity ρ weighting), IR, STIR, and FLAIR with full longitudinal recovery/inversion terms. Fat saturation applies a 90% signal attenuation factor to lipid-labelled voxels. Gadolinium contrast modifies T₁ via the relaxivity relation 1/T₁,post = 1/T₁,pre + r₁[Gd], with r₁ = 4.5 mM⁻¹s⁻¹ at 1.5 T, and models flow-void signal loss in vascular structures.

K-SPACE RECONSTRUCTION PIPELINE

Image formation follows a physically authentic reconstruction chain. A 512×512 complex signal matrix S(x,y) is generated from the tissue map and Bloch-equation signal values, with separate fat and non-fat channels to permit frequency-direction chemical shift displacement (Δx = δf_cs/BW, where δf_cs = 224 Hz × B₀/1.5). The composite image undergoes a 2D radix-2 Cooley–Tukey FFT to produce complex k-space data K(kₓ,kᵧ). All subsequent manipulations occur in the frequency domain:

Matrix truncation. K-space is zeroed outside the user-selected frequency and phase acquisition window (m_freq × m_phase lines centred at DC). The resulting Sinc-convolution in image space produces Gibbs truncation ringing at sharp signal boundaries — the artefact is not simulated; it emerges naturally from the mathematics of finite Fourier sampling.

TSE T₂ modulation. For turbo spin echo acquisitions, each phase-encode line is weighted by exp(−ΔTE(k)/T₂,eff), where ΔTE(k) is the temporal offset from the effective TE based on the line's radial distance from k-space centre and the echo spacing (≈10 ms). This applies a low-pass filter in the phase direction whose width is proportional to ETL, reproducing the characteristic high-spatial-frequency attenuation responsible for TSE blurring.

Parallel imaging. At acceleration factor R > 1, every Rᵗʰ phase-encode line is zeroed in k-space, producing the FOV/R aliased reconstruction visible when GRAPPA reconstruction is disabled. Toggling GRAPPA ON restores fully-sampled k-space with an SNR penalty of √R × g, where the geometry factor g = 1 + 0.15(R−1) models spatially-varying noise amplification from coil sensitivity inversion.

Noise injection. Complex Gaussian noise is added to k-space prior to inverse transform: n(k) ~ ℂ𝒩(0, σₖ²), where σₖ is derived from the voxel-volume SNR model σ ∝ V_vox½ × √NEX × B₀ / (√BW × √R × g). This is physically correct: receiver noise enters in the frequency domain, producing spatially uncorrelated noise after magnitude reconstruction.

K-space spike artefacts. User-placed point impulses are inserted at (kₓ,kᵧ) with Hermitian-conjugate symmetry enforced at (−kₓ,−kᵧ) to maintain real-valued image reconstruction. Each spike produces a sinusoidal modulation pattern across the image at a spatial frequency and orientation determined by its k-space coordinates — reproducing the 'corduroy' or 'zipper' artefact caused by RF interference or electronic faults.

The manipulated k-space undergoes inverse 2D FFT followed by magnitude reconstruction |√(Re² + Im²)|. The result is downsampled to the acquisition matrix, with optional zero-fill interpolation (k-space zero-padding) for display. Window/level adjustment and a transparent anatomical edge overlay (Sobel gradient magnitude extracted from original MR data) are applied as final compositing steps.

ARTEFACT FIDELITY

Every artefact in the Talbot 3000 is a consequence of the acquisition physics, not a post-hoc visual effect. Gibbs ringing arises from k-space truncation. Phase wrap arises from sub-Nyquist spatial sampling. Chemical shift arises from the fat–water frequency offset in the readout gradient. TSE blurring arises from T₂ decay across the echo train. Parallel imaging aliasing arises from phase-encode undersampling. Noise texture arises from complex Gaussian contamination in the frequency domain. The simulator does not contain a single line of code that draws, overlays, or fakes an artefact. They emerge.