Introduction

As intravoxel incoherent motion (IVIM)¹ can provide information about tissue perfusion without the use of contrast agents, there is interest in using it repeatedly to assess longitudinal changes in head and neck (H&N) tumours and organs at risk (OARs) during radiotherapy^2,3. Previous work has shown how different diffusion-weighted MRI models are favoured in different regions of soft tissue sarcomas⁴ and prostate cancer bone metastases⁵. This work introduces a framework for comparing IVIM and apparent diffusion coefficient (ADC) models, with the hypothesis that IVIM-favoured voxels reflect perfused regions. Technical validation⁶ is performed through simulations, phantom experiments, and repeated pre-treatment scans of patients with H&N cancer.

Methods

Simulations: Signals were generated according to IVIM (S = S₀[fexp(-bD*) + (1-f)exp(-bD)] ) and ADC ( S = S₀exp(-bADC) ) models, where f is the perfusion signal fraction, D* is the pseudo-diffusion coefficient, D is the tissue diffusion coefficient; two sets of model parameter values were used, taken from literature values for the parotid glands². b-values of 0, 20, 40, 60, 80, 100, 150, 200, 300, 400, 500, 800 s/mm² were used, matching those used experimentally (see below). Noise was added to the signals, with all resulting noisy signals fitted using both IVIM and ADC expressions. Fits were compared using the corrected Akaike information criterion (AICc), taking the model with the lowest AICc as the preferred model. The ability to correctly identify the model from which the data were generated was evaluated for synthetic ‘tissues’ with a range of ground truth proportions of IVIM-favoured voxels, p_IVIM.

Phantom experiments: An ice-water phantom (High Precision Devices) with 13 ADC values (0.1 – 1.1 μm²/ms) was scanned on a 1.5 T Philips Ingenia using a single-shot EPI diffusion sequence (12 b-values as above, TR = 3000 ms, TE = 65 ms, voxel size = 3x3x5 mm³), and the model comparison framework was applied. The ADC model is predicted to be preferred throughout, as the phantom has no perfusion component.

Patient scans: Eight patients with H&N tumours were scanned using the same scanner and sequence as the phantom experiments. Patients were recruited to the imaging sub-study of the TORPEdO trial⁶, and provided written informed consent. Patients had two repeated scans pre-treatment (5±2 days apart) to assess imaging biomarker repeatability, before starting radiotherapy. Patients were scanned in their radiotherapy treatment planning position, with a thermoplastic shell. The model comparison framework was applied to the pre-treatment datasets, to evaluate p_IVIM in the parotid glands.

Results

Simulations: Figure 1 plots estimated values of the proportion of IVIM-favoured voxels against the ground truth. At realistic noise levels there is a tendency to misclassify IVIM voxels as ADC voxels, resulting in an underestimation of p_IVIM; the magnitude of the underestimation tends to increase as p_IVIM increases. The degree of underestimation also depends on model parameter values, with higher D* and f values aiding the distinction between ADC and IVIM voxels (Figure 1, comparing left and right plots).

Phantom experiments: Figure 2 shows model preference maps for three central slices of the phantom, along with b = 0 s/mm² images and ADC maps. As expected, the ADC model (blue) tends to be preferred throughout the 13 vials. For the vials with the lowest ADCs (<~0.4 μm²/ms), a number of voxels show preference for the IVIM model. Spurious IVIM-favoured voxels are also observed at vial edges and in ice water surrounding the vials.

Patient scans: Figure 3 shows plots of signal against b-value for three voxels in the parotid gland, illustrating that ADC and IVIM models may be preferred in different regions. Figure 4 shows model preference maps overlaid on a b = 0 s/mm² image, again showing the presence of ADC-favoured and IVIM-favoured voxels across both parotids. Figure 5 plots p_IVIM for both pre-treatment scans. Across all patients and visits, p_IVIM = 0.79 ± 0.05, with a within-subject coefficient of variation of 5.4%.

Discussion

Taken together, the simulations and phantom data highlight different factors which can lead to bias in quantifying the proportion of voxels where IVIM is preferred. Low SNR can lead to an underestimation of p_IVIM, with genuine IVIM voxels misclassified as favouring ADC, likely due to noisy data being unable to support the increased complexity of the IVIM model. The phantom data indicate that in low ADC regions, voxels may be incorrectly identified as favouring the IVIM model; however, for clinically-relevant ADC values (>0.4 μm²/ms), areas genuinely lacking a perfusion component are reliably identified as such. In the parotid glands there is evidence of both models being preferred, indicating that applying a single model to all voxels may not be appropriate. Despite the potential biases indicated by the simulations and phantom data, the proportion of voxels favouring IVIM, and hypothesised to include a perfusion component, exhibits good repeatability in vivo.

Conclusion

The combination of simulations and phantom data highlights how SNR and diffusion properties can bias p_IVIM, while patient scans show it is repeatable in the parotid glands. This technical validation aids interpretation of attempts to identify perfused tissue regions through ADC and IVIM model comparison. Future work will aim to quantify uncertainties in the model comparison framework using the magnitude of the difference in AICc values.

Acknowledgements

This work was supported by funding from Cancer Research UK (CRUK/18/010) and The Taylor Family Foundation. The authors acknowledge The Institute of Cancer Research as the sponsor of the TORPEdO trial.

References

[1] Le Bihan. NeuroImage, 2019;187:56–67. [2] Zhou et al. BMC Cancer, 2016;16:865. [3] Marzi et al. Radiat Oncol, 2018;13:189. [4] Manikis et al. Eur J Radiol, 2021;138:109660. [5] Reischauer et al. Sci Rep, 2017;7:5809. [6] O'Connor et al. Nat Rev Clin Oncol, 2017;14:169-186. [7] Price et al. Clin Oncol, 2020;32:84-88.