T2 in MRI: Relaxation, Mapping, and Verification

• T2 is the transverse relaxation time in MRI, modeled by exponential decay to quantify tissue properties and guide imaging protocols.
• Recent advances integrate model-based and learning-based methods to accelerate T2 mapping, improve quantification accuracy, and enable advanced radiomics.
• Beyond MRI, T2 also refers to an open-source tool for temporal property verification, underscoring its context-dependent technical significance.

In the cited MRI literature, T2 denotes the transverse relaxation time, estimated from signal models such as $M_\perp(t)=M_0 e^{-t/T_2}$ or $\hat{X}_{TE}=\mathcal{M}_0 e^{-TE/T2}$, and used in T2-weighted imaging, T2 mapping, T2 preparation, and quantitative relaxometry (Zhang et al., 16 Dec 2025, Lajous et al., 2020). Across recent arXiv work, T2 functions both as a physical parameter to be quantified and as an imaging contrast for downstream tasks such as detection, segmentation, radiomics, and cancer localization. In a distinct computer-science usage, T2 is also the name of an open-source tool for temporal property verification (Brockschmidt et al., 2015).

1. Transverse relaxation and formal signal models

A simplified Bloch-equation formulation for T2 decay writes

$$\frac{d M_\perp(t)}{dt} + \frac{M_\perp(t)}{T_2} = 0, \qquad M_\perp(t) = M_0 e^{-t/T_2},$$

and this model is used directly in a Physics-Informed Neural Network for cardiac T2 quantification (Zhang et al., 16 Dec 2025). In multi-echo mapping work, the same quantity is expressed voxelwise as

$\hat{X}_{TE} = \mathcal{M}_0 \, e^{-TE / T2},$

with $\mathcal{M}_0$ denoting equilibrium magnetization and $T2$ the transverse relaxation time (Lajous et al., 2020). These formulations underlie both conventional nonlinear fitting and more recent model-based or learning-based estimation pipelines.

Recent simulation work places T2 within the broader decomposition of effective transverse relaxation,

$\frac{1}{T_2^*} = \frac{1}{T_2} + \frac{1}{T_2'},$

where $T_2$ is the irreversible component and $T_2'$ the reversible component induced by local magnetic-field inhomogeneities (Takeshima, 27 Jan 2026). That paper proposes a linear phase model for simulating the Lorentzian behavior of $T_2'$ directly, avoiding the need for $\hat{X}_{TE}=\mathcal{M}_0 e^{-TE/T2}$0 isochromats per point and recovering $\hat{X}_{TE}=\mathcal{M}_0 e^{-TE/T2}$1 with computational times only $\hat{X}_{TE}=\mathcal{M}_0 e^{-TE/T2}$2 to $\hat{X}_{TE}=\mathcal{M}_0 e^{-TE/T2}$3 times longer than simulations without $\hat{X}_{TE}=\mathcal{M}_0 e^{-TE/T2}$4 (Takeshima, 27 Jan 2026).

The classical assumption that T2 relaxation during an RF pulse is negligible is sequence-dependent. For ultrashort echo time and zero echo time imaging, the relevant regime is one in which $\hat{X}_{TE}=\mathcal{M}_0 e^{-TE/T2}$5 is of the order of, or shorter than, the RF pulse duration, so relaxation during the pulse materially affects excitation efficiency, slice profile, image contrast, and SNR (Larson, 2023). This establishes T2 not only as a post-acquisition parameter but also as a determinant of pulse-design behavior.

2. Quantification and estimation paradigms

Conventional T2 quantification remains anchored in multi-echo acquisition and voxelwise fitting, but recent work emphasizes how the estimation problem changes under motion, low field, undersampling, or limited echo sampling. In fetal-brain-oriented relaxometry, super-resolution reconstruction from conventional T2-weighted fast spin echo acquisitions was shown to enable 3D isotropic T2 mapping from low-resolution 2D HASTE stacks. With six echo times, total acquisition time was under 9 minutes, relative error was below $\hat{X}_{TE}=\mathcal{M}_0 e^{-TE/T2}$6 over the $\hat{X}_{TE}=\mathcal{M}_0 e^{-TE/T2}$7 ms range, and repeatability for the super-resolution method was reported as mean CV below $\hat{X}_{TE}=\mathcal{M}_0 e^{-TE/T2}$8 (Lajous et al., 2020). That work explicitly identified six echo times as the optimal tested configuration.

Low-field implementation introduces a separate estimation regime. At $\hat{X}_{TE}=\mathcal{M}_0 e^{-TE/T2}$9 T, mono-exponential fitting under Gaussian--Rician noise in vitro yielded deviations below $$\frac{d M_\perp(t)}{dt} + \frac{M_\perp(t)}{T_2} = 0, \qquad M_\perp(t) = M_0 e^{-t/T_2},$$0 from spectrometer references, and the in vivo protocol produced mean T2 times of $$\frac{d M_\perp(t)}{dt} + \frac{M_\perp(t)}{T_2} = 0, \qquad M_\perp(t) = M_0 e^{-t/T_2},$$1 ms in white matter and $$\frac{d M_\perp(t)}{dt} + \frac{M_\perp(t)}{T_2} = 0, \qquad M_\perp(t) = M_0 e^{-t/T_2},$$2 ms in grey matter with a $$\frac{d M_\perp(t)}{dt} + \frac{M_\perp(t)}{T_2} = 0, \qquad M_\perp(t) = M_0 e^{-t/T_2},$$3-minute acquisition (Roulet et al., 22 Oct 2025). Inter-subject CoV was $$\frac{d M_\perp(t)}{dt} + \frac{M_\perp(t)}{T_2} = 0, \qquad M_\perp(t) = M_0 e^{-t/T_2},$$4 for white matter and $$\frac{d M_\perp(t)}{dt} + \frac{M_\perp(t)}{T_2} = 0, \qquad M_\perp(t) = M_0 e^{-t/T_2},$$5 for grey matter, reported as comparable to $$\frac{d M_\perp(t)}{dt} + \frac{M_\perp(t)}{T_2} = 0, \qquad M_\perp(t) = M_0 e^{-t/T_2},$$6 T (Roulet et al., 22 Oct 2025). The same study investigated Gaussian, Rician, and Gaussian--Rician models and adopted Gaussian noise-based fitting as the practical mapping choice for the main workflow (Roulet et al., 22 Oct 2025).

Model-based quantification has also acquired explicit theoretical guarantees. A cardiac MRI study embedded the Bloch equation in the loss of a PINN and derived upper bounds for both T2 estimation error and the generalization error of the Bloch-equation solution, without requiring a pre-defined training database (Zhang et al., 16 Dec 2025). Validation spanned a numerical cardiac model, a water phantom, and $$\frac{d M_\perp(t)}{dt} + \frac{M_\perp(t)}{T_2} = 0, \qquad M_\perp(t) = M_0 e^{-t/T_2},$$7 acute myocardial infarction patients; in infarct regions, the theoretical error bound was reported as $$\frac{d M_\perp(t)}{dt} + \frac{M_\perp(t)}{T_2} = 0, \qquad M_\perp(t) = M_0 e^{-t/T_2},$$8, corresponding to T2 estimation errors of $$\frac{d M_\perp(t)}{dt} + \frac{M_\perp(t)}{T_2} = 0, \qquad M_\perp(t) = M_0 e^{-t/T_2},$$9 ms in typical lesion ranges (Zhang et al., 16 Dec 2025). A notable result of that study was that $\hat{X}_{TE} = \mathcal{M}_0 \, e^{-TE / T2},$0 closely tracked empirical error in synthetic and phantom experiments, with mean ratio $\hat{X}_{TE} = \mathcal{M}_0 \, e^{-TE / T2},$1 and standard deviation $\hat{X}_{TE} = \mathcal{M}_0 \, e^{-TE / T2},$2 across tested noise and sampling conditions (Zhang et al., 16 Dec 2025).

3. Accelerated acquisition and reconstruction

A dominant theme in recent T2 research is the replacement of long, separate relaxometry acquisitions with accelerated, joint, or subspace-based protocols. These methods combine tailored encoding, structured reconstruction, and quantitative fitting so that high-resolution T2 maps are produced with clinically shorter scan times.

Approach	Reported capability	Representative output
Wave-Shuffling (Iyer et al., 2021)	$\hat{X}_{TE} = \mathcal{M}_0 \, e^{-TE / T2},$3 mm-isotropic, time-resolved structural imaging; $\hat{X}_{TE} = \mathcal{M}_0 \, e^{-TE / T2},$4 seconds for 3D-FSE brain imaging and $\hat{X}_{TE} = \mathcal{M}_0 \, e^{-TE / T2},$5 seconds for MPRAGE	Time-series of T$\hat{X}_{TE} = \mathcal{M}_0 \, e^{-TE / T2},$6- or T$\hat{X}_{TE} = \mathcal{M}_0 \, e^{-TE / T2},$7-weighted images, up to $\hat{X}_{TE} = \mathcal{M}_0 \, e^{-TE / T2},$8 per scan
3D-QALAS with Wave-CAIPI (Cho et al., 2022)	Whole-brain quantitative mapping at $\hat{X}_{TE} = \mathcal{M}_0 \, e^{-TE / T2},$9 mm$\mathcal{M}_0$0 isotropic voxels in $\mathcal{M}_0$1 minutes at $\mathcal{M}_0$2	Simultaneous T1, T2, and PD maps
T2-BUDA-gSlider (Cao et al., 2019)	Distortion-free whole-brain T2 mapping in $\mathcal{M}_0$3 min at $\mathcal{M}_0$4 mm$\mathcal{M}_0$5 isotropic resolution	T2 values consistent with gold standard spin-echo acquisition
BUDA-SAGE with Slider (Zhang et al., 2021)	Whole-brain, distortion-free T2 and T2$\mathcal{M}_0$6 mapping at $\mathcal{M}_0$7 mm$\mathcal{M}_0$8 isotropic resolution in $\mathcal{M}_0$9 seconds	Quantitative T2, T2$T2$0, QSM, and para-/dia-magnetic susceptibility maps
RADTSE (Keerthivasan et al., 2019)	Accurate T2 maps from a single acquisition using highly undersampled radial turbo spin echo data	Whole abdominal coverage in up to $T2$1 breath-holds

These protocols differ in mechanism. Wave-Shuffling augments the Shuffling subspace model with wave-encoding, combining a wave point spread function $T2$2, coil sensitivities estimated by ESPIRiT, ShflPSF auto-calibration, and FISTA with locally low-rank regularization (Iyer et al., 2021). T2-BUDA-gSlider couples RF-encoded multi-slab spin-echo EPI, blip-up/down phase encoding, structured low-rank reconstruction, and a Bloch-simulated subspace model to obtain distortion-free whole-brain T2 mapping in $T2$3 minutes (Cao et al., 2019). BUDA-SAGE adds self-supervised MR-Self2Self denoising and Slider super-resolution, then uses Bloch dictionary matching to obtain T2 and T2$T2$4 maps from reconstructed echoes (Zhang et al., 2021).

Radial methods remain relevant when motion robustness and abdominal coverage are primary constraints. RADTSE uses a radial TSE pulse sequence in which each spoke passes through k-space center, so TE images are co-registered and robust to motion; accurate T2 mapping depends on physical modeling of stimulated and indirect echoes, with SEPG dictionary fitting outperforming mono-exponential fitting in phantom measurements (Keerthivasan et al., 2019).

4. Learning-based and radiomic analysis on T2 data

T2-weighted images and T2 maps increasingly serve as direct inputs to detection, segmentation, localization, and radiomic pipelines. In abdominal T2 MRI, a transformer-based detector used the DEtection TRansformer architecture with a ResNet-50 backbone pretrained on MS COCO, positional encoding, encoder-decoder self-attention, bipartite set matching, and focal loss weighted by $T2$5 for classification cost (Mathai et al., 2021). On $T2$6 MRI scans with $T2$7 annotated lymph nodes, the method achieved mAP $T2$8, precision $T2$9, and sensitivity $\frac{1}{T_2^*} = \frac{1}{T_2} + \frac{1}{T_2'},$0 at $\frac{1}{T_2^*} = \frac{1}{T_2} + \frac{1}{T_2'},$1 false positives per image; smaller nodes with $\frac{1}{T_2^*} = \frac{1}{T_2} + \frac{1}{T_2'},$2 mm remained harder, with mAP $\frac{1}{T_2^*} = \frac{1}{T_2} + \frac{1}{T_2'},$3 and sensitivity $\frac{1}{T_2^*} = \frac{1}{T_2} + \frac{1}{T_2'},$4 at $\frac{1}{T_2^*} = \frac{1}{T_2} + \frac{1}{T_2'},$5 FP/image (Mathai et al., 2021). Weighted Boxes Fusion of the five lowest-validation-loss models was used to reduce false positives (Mathai et al., 2021).

In cardiac MRI, SQNet formalized simultaneous myocardium segmentation and T2 quantification as a dual-task network combining Transformer and CNN components through a tight coupling module and a T2-refine fusion decoder (Zhou et al., 2024). Reported segmentation Dice scores were $\frac{1}{T_2^*} = \frac{1}{T_2} + \frac{1}{T_2'},$6 for healthy controls and acute myocardial infarction patients, compared with $\frac{1}{T_2^*} = \frac{1}{T_2} + \frac{1}{T_2'},$7 for the cited state-of-the-art baseline, while Pearson coefficients for T2 quantification were $\frac{1}{T_2^*} = \frac{1}{T_2} + \frac{1}{T_2'},$8 for healthy control and AMI cohorts (Zhou et al., 2024). Radiologist scores were $\frac{1}{T_2^*} = \frac{1}{T_2} + \frac{1}{T_2'},$9 for segmentation and $T_2$0 for T2 quantification, exceeding the cited comparison methods (Zhou et al., 2024).

A distinct learning setup appears in prostate MRI, where only T2-weighted images are available at inference but DWI is treated as a latent modality during training. The proposed expectation-maximization framework uses a flow matching-based generative model in the E-step to approximate the latent DWI posterior and a cancer localizer in the M-step to maximize expected likelihood of cancer presence (Yi et al., 1 Apr 2026). On internal and external cohorts totaling $T_2$1 prostate cancer patients with histopathology-verified labels, the reported T2-only method improved patient-level F1 score by $T_2$2 and zone-level QWK by $T_2$3 over the T2w+DWI baseline (Yi et al., 1 Apr 2026). This suggests that, in some formulations, T2-weighted input can be treated not merely as a reduced-modality compromise but as the anchor modality in a privileged-learning framework.

Radiomic analysis exposes a different limitation of scalar averaging. In cadaveric knee stress MRI, mean T2 values showed no significant dependency on pressure or meniscectomy, whereas the radiomic texture parameter variance detected loading-induced textural changes in medial femoral cartilage with $T_2$4 for UL vs. LN, $T_2$5 for UL vs. LV, and $T_2$6 for LN vs. LV (Lemainque et al., 21 Aug 2025). The paper concluded that T2-based radiomic features were more sensitive than mean T2 for assessing cartilage response to loading (Lemainque et al., 21 Aug 2025).

5. Sequence design, RF preparation, and confounding mechanisms

T2 is not only measured; it is also engineered into acquisition modules. In cardiac MRI at $T_2$7 T, a phaser adiabatic T2-Prep replaced the tip-down RF pulse of a conventional adiabatic T2-preparation module with a custom-designed excitation pulse that induces a phase difference between water and fat, producing simultaneous water T2 preparation and fat suppression without an additional fat-suppression module (Arn et al., 2020). Numerical simulations predicted a fat-suppression bandwidth increase from $T_2$8 Hz to $T_2$9 Hz and a wider effective B1 range for fat suppression, $T_2'$0 rather than $T_2'$1, while preserving water T2-preparation capability (Arn et al., 2020). In vivo, right coronary artery vessel sharpness increased by $T_2'$2, with minimal SAR increase of $T_2'$3 (Arn et al., 2020).

The interaction of T2 with RF excitation becomes especially important for short-T2 species. When $T_2'$4 or $T_2'$5, the full Bloch equations with relaxation must be considered during the pulse, and the literature explicitly links this regime to degraded excitation efficiency and altered slice profiles in UTE and ZTE imaging (Larson, 2023). Sequence design in this regime therefore emphasizes short, high-bandwidth RF pulses and, where needed, low-bandwidth pulses for selective suppression or inversion of long-T2 species (Larson, 2023).

T2 can also act as a confound in ostensibly non-relaxometric measurements. A 2023 letter argues that apparent diffusion coefficient values contain information from both diffusion and T2 relaxation, because $T_2'$6 at $T_2'$7 is T2-weighted and high-$T_2'$8 images may be noise-dominated (Wang, 2023). The paper supports that argument with examples involving uterine myometrium tumors, spleen versus liver, gallbladder, and parotid tumors, and states that, in extreme cases, ADC may be devoid of diffusion information (Wang, 2023). This remains a pointed interpretive claim rather than a settled consensus in the supplied corpus.

6. Clinical and experimental domains

The application range of T2-based methods is broad. In fetal-brain imaging, super-resolution T2 mapping was motivated by the impracticality of long relaxometry protocols in utero and by the routine use of 2D low-resolution HASTE sequences to mitigate unpredictable fetal motion (Lajous et al., 2020). In low-field neuroimaging, $T_2'$9 T HASTE-based T2 mapping was proposed to improve accessibility and provide quantitative biomarkers of brain development, with the first normative T2 values for healthy adult brains at that field strength reported as $T_2'$0 ms in white matter and $T_2'$1 ms in grey matter (Roulet et al., 22 Oct 2025).

Cardiac MRI provides both quantitative and preparatory uses of T2. PINN-based T2 quantification was validated in $T_2'$2 acute myocardial infarction patients and used theoretical error bounds to estimate quantitative accuracy without a gold standard (Zhang et al., 16 Dec 2025). SQNet addressed simultaneous myocardium segmentation and T2 mapping in healthy controls and AMI patients (Zhou et al., 2024), while phaser adiabatic T2-Prep targeted robust fat suppression and vessel conspicuity in coronary MR angiography at $T_2'$3 T (Arn et al., 2020).

Abdominal and oncologic applications rely on T2 either as a detection substrate or as a quantitative target. DETR-based lymph-node localization addressed challenging T2 MRI scans acquired by different scanners and exam protocols, with sensitivity $T_2'$4 at $T_2'$5 FP/image (Mathai et al., 2021). RADTSE focused on quantitative T2 estimation in abdominal imaging from a single highly undersampled acquisition (Keerthivasan et al., 2019). Prostate MRI work pursued T2-only cancer localization from histopathology labels by exploiting DWI as a privileged latent modality during training (Yi et al., 1 Apr 2026). Cartilage stress MRI, by contrast, showed that loading effects may be better captured by T2 texture features than by mean T2 itself (Lemainque et al., 21 Aug 2025).

A plausible implication is that T2 now occupies three roles simultaneously: a tissue parameter, a pulse-sequence design target, and a computational representation for downstream inference. The supplied literature supports all three roles directly.

7. T2 as a temporal property verification tool

Outside MRI, T2 is the name of an open-source tool for verifying temporal properties of potentially infinite-state programs (Brockschmidt et al., 2015). Originating from the TERMINATOR project, it supports safety, termination, non-termination, CTL, Fair-CTL, CTL$T_2'$6, and LTL, and accepts input in a native format and in C through the LLVM compiler framework (Brockschmidt et al., 2015). Its architecture translates verification tasks into safety queries over instrumented programs, uses Horn-clause encodings and safety engines such as Impact and Z3 backends, and synthesizes lexicographic ranking functions or recurrent sets for termination and non-termination reasoning (Brockschmidt et al., 2015).

This second usage is unrelated to MRI physics, but it is technically significant because it shows that “T2” is field-dependent shorthand rather than a globally unique scientific term. In arXiv practice, contextual disambiguation is therefore essential: in medical imaging it almost always denotes transverse relaxation, whereas in formal verification it denotes a specific software system (Brockschmidt et al., 2015).