Uterine cervical cancer poses a significant global health challenge, particularly in developing regions where access to preventive measures and screening are often limited. Locally advanced cases frequently necessitate concurrent chemoradiation, aiming to achieve optimal local control and minimize the risk of recurrence (1).

There is a need for validated imaging biomarkers (2) to assess the early response of cervical tumours to therapy. Diffusion-weighted (DW) MRI is a functional imaging technique which provides various quantitative biomarkers that have potential to evaluate tumour response to therapy (3). Modelling the DW signal as a mono-exponential decay with b-value yields the apparent diffusion coefficient (ADC), the simplest quantitative biomarker to measure from DW-MRI, with several studies reporting values in cervical tumours (3–7).

Intra-voxel incoherent motion (IVIM) is a DW-MRI method that uses a bi-exponential decay model to separate the effects of tissue diffusion and capillary blood flow; as such, it provides more specific information about tumour microstructure and microvasculature than ADC, yielding parameters such as the tissue diffusion coefficient, D, and the perfusion signal fraction, f (8). IVIM does not require gadolinium-based contrast agents, and parameters have distinguished cervical tumours from non-malignant uterine tissues (9), and distinguished between cervical tumour histological subtypes and/or grades (10–12). Several studies have investigated the ability of IVIM biomarkers to predict and assess cervical tumour treatment response (13–17). Recent studies have also shown the utility of IVIM in predicting parametrial invasion (18), and shown that combined IVIM and FDG PET can identify lymphovascular invasion (19) and treatment resistance (20) in cervical cancer.

IVIM biomarkers require validation if they are to be translated into clinical practice (2). In particular, the IVIM model may not be applicable in all tumour regions (12) and model suitability may vary over the course of therapy. Previous work has shown spatial and temporal variation in model suitability for non-IVIM DW-MRI models (21), but this type of analysis has not yet been performed in IVIM studies of cervical cancer.

As IVIM requires longer scan times and more complex model fitting than ADC, IVIM biomarkers must provide additional utility over ADC, and potential bias when using the simpler ADC biomarker must be understood. In particular, several studies reporting ADC in cervical tumours use b = 0 s/mm² as the lowest b-value (3–7), which is expected to make such ADCs sensitive to perfusion effects (22). As well as increasing ADC values, this may impact the evaluation of treatment-induced ADC changes if treatment affects both tumour tissue and vasculature.

Here we employ a model comparison framework to assess the spatial and temporal variability in suitability of the IVIM model, investigating both healthy uterine tissue and uterine cervical tumours (23). IVIM repeatability is evaluated through test-retest scanning of healthy volunteers. The sensitivity of IVIM to therapy-induced changes, and the impact of perfusion on ADC measurements, is assessed in patients with cervical cancer.

Research ethics committee approval was obtained. Fully informed written consent was obtained from all participants. Power calculations were not performed for this feasibility study.

Six healthy volunteers (mean ± standard deviation [s.d.] age: 26 ± 1 years; all pre-menopausal) were scanned in two sessions a median of 8 days apart (range: 7–70 days). Eight patients (age: 47 ± 19 years; four pre-menopausal; Supplementary Table S1) were scanned, with pre-treatment scans a median of 6 days before starting treatment (range: 3–27 days). Six patients were scanned at all three time points, and two missed the week 3 scan due to scheduling difficulties during COVID-19 restrictions.

This work contributes to DW-MRI biomarker validation by showing that both IVIM- and ADC-favoured voxels exist within tumours; IVIM parameters are repeatable in healthy tissue and differ from those in tumours; D is sensitive to treatment-induced tumour changes; and that ADC values can suffer from a perfusion-dependent bias. The model comparison has shown that not all tumour tissue is best described by the IVIM model, p_IVIM has been introduced as a novel biomarker, and the bias in ADCs calculated with b = 0 s/mm² has been shown to depend on f.

The model preference maps reflect expected trends based on the physiological characteristics of different tissues. In tumours, where intra- and inter-lesion microenvironment heterogeneity is expected, spatial variation and large inter-patient variability in model preference was observed. This shows that one MR signal model may not be applicable in all tumour voxels at all time points, highlighting that statistical model comparison should be considered when validating imaging biomarkers. Note that the accuracy of model comparison approaches has been shown to decrease with lower signal-to-noise ratios (SNR) (21), which may lead to an underestimation of p_IVIM. This is an inherent limitation, with simpler models being increasingly preferred as data becomes noisier, which in this context would lead to ADC being preferred over IVIM. Future work could incorporate SNR into the model comparison, and incorporate the quantitative difference between models’ AIC_c values (24), to further assess the bias and precision of p_IVIM. Nevertheless, the observed qualitative consistency between the model preference maps and known tissue characteristics provides evidence that the model comparison provides biologically meaningful information.

D, f, and p_IVIM exhibited good repeatability in the cervix of healthy volunteers, with D being the most repeatable. D was also the most repeatable in the uterine body, though for all parameters repeatability was poorer in the uterine body compared to the cervix. This may be partly due to the uterine body being more susceptible to inter-b-value misalignment, stemming from bladder filling and bowel motion. These effects should be more pronounced in the healthy volunteers, as Buscopan was not administered to help suppress bowel motion. Parameter estimates may also be affected by partial voluming with the endometrium and endocervical canal. For both ROIs, some variability may also be due to the repeated scans being performed at different times in the subjects’ menstrual cycle, which influences diffusion in uterine and cervical tissue (34, 35). Repeatability would therefore be expected to improve if repeat scans were performed in the same menstrual phase, with values reported here reflecting a worst-case scenario. Nevertheless, wCV for IVIM biomarkers were similar to estimates of ADC repeatability (36, 37) and lower than those derived from DCE-MRI (38, 39).

D and f differed significantly between healthy volunteers’ cervix tissue and patients’ cervical tumours, consistent with previous reports (9). Absolute values also agree well (9), though differences in sequence parameters, healthy volunteers’ age, and patients’ disease stage confound a direct comparison between studies. In the present study, the healthy volunteer cohort was younger than the patient group, and all healthy volunteers were pre-menopausal, which may contribute to the significant difference in uterine body D (40). As healthy cervix D has been reported to not vary with age (41), and significantly lower D has been reported in cervix tumours compared with age-matched healthy volunteers’ normal cervix (9), we expect age differences to be a minor factor in our comparison of healthy volunteers’ cervix and patients’ cervical tumours.D was the only parameter which showed sensitivity to treatment-induced changes, with an increase in D consistent with IVIM findings (13), and with ADC increasing following therapy (42). An increase in D may reflect a loss of cell membrane integrity and/or decreases in cell density as a result of chemoradiation (13), with some evidence that on-treatment D changes differ between responders and non-responders (14). Increases in f have been reported previously (13), but this was not observed in the present study. This may be due to the small patient numbers and greater variability of f compared with D, as indicated by the higher wCV in healthy volunteers. Changes in f are expected to reflect chemoradiation-induced changes in tumour microvasculature, with higher f values, suggesting higher perfusion, reported for responders than non-responders (14). Tumour p_IVIM also did not change significantly throughout treatment; as for f, this may be due to small patient numbers and poorer repeatability than D, based on healthy volunteer wCVs. p_IVIM is introduced here as a novel biomarker, with no previous reports in the literature, but we hypothesise that p_IVIM may have sensitivity to treatment if this induces changes in the proportions of distinct tumour microenvironments. While menopausal status has been shown to impact absolute values of DW-MRI tumour biomarkers (43), we are not aware of literature showing it has an impact on treatment-related changes, so do not consider this a confounding factor here.

As tumour volume tended to decrease throughout treatment, contouring was challenging at week 5; this is consistent with a previously reported increase in inter-observer variability in cervical tumour delineations post-treatment relative to baseline (17). Future studies may benefit from an earlier final time point. To complement the healthy volunteer repeatability reported in this work, future studies evaluating tumour biomarker repeatability are warranted, and could help determine biomarker sensitivity to treatment-induced changes (44). A ‘coffee-break’ assessment of short-term repeatability (45) may be beneficial here, removing the need for additional patient visits.

The AIC_c model comparison showed that the IVIM model is not favoured in all tumour voxels. This is expected to reflect a combination of noise, which can result in the simpler ADC model being favoured, and genuine spatial variation in the tumour microenvironment (46). ADC-favoured voxels are hypothesised to reflect regions without a measurable perfusion component, which could be due to the presence of necrosis or oedema, or areas of viable tumour with a very low vascular volume fraction. The model comparison shows that one signal model may not be applicable in all tumour voxels at all time points, highlighting that statistical model comparison should be considered when validating imaging biomarkers. Moreover, not accounting for model suitability can lead to a bias in summary model parameters, as shown here with a tendency for lower median f when including all tumour voxels, compared with only including voxels where IVIM is favoured. In addition, model suitability can impact diffusivities in voxels favoured by the ADC model, as shown here with a tendency for ADC_b0 to yield higher diffusivities than D. The optimal approach to reporting summary statistics when there is spatial variability in model suitability requires further investigation; one option would be to only calculate parameter summary statistics from voxels where that parameter’s associated model has been selected.

For clinical translation, the fitting and model comparison pipeline would need to be incorporated into scanner and/or dedicated post-processing software. ADC maps are routinely generated on scanners, and this processing could be extended to IVIM parameter maps, acknowledging the added complexity of fitting a bi-exponential model (29). In terms of data acquisition, IVIM scan times will be longer than for ADC, though with further optimisation of b-values this time could be reduced (29). Overcoming these technical and practical challenges would aid the translation of IVIM-based biomarkers into clinical practice; for example as predictors of treatment response, as spatial maps to include in biological-image guided adaptive radiotherapy planning (47), or as early indicators of treatment response to guide subsequent treatment approaches. Such translation is supported by existing studies showing associations between pre-treatment IVIM biomarkers and RECIST-based treatment response (17, 48), correlations between IVIM biomarkers and a histology-derived hypoxia biomarker (49), and therapy-induced IVIM biomarker changes differing between responders and non-responders (13, 15).

Comparing ADC_b150 and ADC_b0 highlighted the impact of b-value selection on ADC, with perfusion effects increasing ADC when low b-values are included in the fit. This is especially relevant for cervical tumours, as several studies (3–7) include b = 0 s/mm² and therefore introduce a bias in ADC. Importantly, the magnitude of this bias depends on the perfusion fraction, f, and if f changes throughout treatment, the ADC bias will change too. As such, it is recommended to not include b = 0 s/mm² when evaluating cervical tumour ADC, to reduce the sensitivity to perfusion effects. Here, a pragmatic approach was taken to use a single b-value threshold of 150 s/mm², but the degree to which this suppresses perfusion effects will depend on the underlying perfusion characteristics. This is a limitation of using a single threshold, and for tissue with significant and/or changing perfusion characteristics, it may be more appropriate to model this explicitly using IVIM. This is supported by our results showing that while ADC_b150 has a weaker correlation with f than ADC_b0, a correlation does still exist; moreover, use of D removes this correlation.

The main limitation of this study is the small sample size, for both healthy volunteers and patients. The small numbers reflect the exploratory nature of the trial from which these data come, and the results can be used to power follow-on studies. Further limitations are the lack of control for subjects’ menstrual cycles, and patient choice in having Buscopan to counteract bowel motion, which may be sources of variability in the reported biomarkers. Also, patients and healthy volunteers were not aged-matched, tumour repeatability was not assessed. The single-centre nature of the study limits the extent of technical validation that the study provides, and the study does not address direct biological validation as correlations between imaging and histology (e.g. cell and vessel densities) were not investigated. As the model comparison framework is based on IVIM and ADC model equations, we envisage it being readily applied to PGSE-based DW-MRI data from other vendors and/or field strengths. Future work can therefore extend the technical, biological, and clinical validation of IVIM biomarkers through multi-centre repeatability and reproducibility studies, imaging-histology correlations, and relating imaging biomarkers to outcomes.

Taken together, this study further advances the validation of DW-MRI biomarkers in cervical cancer. The model comparison demonstrated that the IVIM model is not favoured in all tumour voxels, indicating the presence of qualitatively different tumour microenvironments, and further demonstrating that model suitability should be evaluated as part of imaging biomarker validation. IVIM biomarkers were repeatable in healthy tissue, with D sensitive to early therapy-induced changes in tumours. In addition, ADC calculated from low b-values suffers from a perfusion-dependent bias.