Our results suggest that DWI, and specifically 90th percentile ADC value, has clinical relevance in risk-stratifying patients undergoing ISBT for bulky cervical cancer. We also explored the potential for identifying a 90th percentile ADC threshold that can differentiate patients based on likelihood of complete response to brachytherapy.

While multiple studies in recent years have explored the utility of ADC values in predicting outcomes for cervical cancer in general [24, 25, 30,31,32], little attention has been given to the subset of patients who require ISBT after EBRT. These patients represent a high-risk population and thus would greatly benefit from improved prognostication. Our work suggests that the prognostic utility of ADC values for cervical cancer treated with EBRT can be extended to those tumors requiring ISBT.

Despite mixed results as to whether lower ADC values portend a better or worse prognosis [23, 33], the literature in recent years has generally supported an association between higher ADC values and better outcomes [24, 25, 31, 32, 34,35,36]. Our results are in line with this finding. Figure 3A illustrates an important bimodal distribution: while lower 90th percentile ADC values are spread between the CMR and non-CMR groups, the cohort of patients with high 90th percentile ADC values ( ≥ ~ 2.5 × 10− 3 mm2/s) almost universally achieved a CMR. The upward spread of ADC values in these DWI images likely represents tumor necrosis after EBRT (e.g., Fig. 1D, Patient B, black arrow in the ADC map), portending a favorable response to ISBT [34].

Despite consensus on the direction of this association, agreement has not yet been reached as to ADC cutoff values that could be used to risk-stratify cervical cancer patients. Ho, et al. found a 90th percentile pretreatment ADC cutoff value of 1.917 × 10− 3 mm2/s, with lower values associated with worsened outcomes [25], while our data suggests a 90th percentile ADC cutoff value of 2.10 × 10− 3 mm2/s. Two notable distinctions are that in our analysis our scans were all acquired following EBRT and prior to BT (rather than at baseline) and our focus was on metabolic response, given our population of bulky local disease, rather than survival or progression-free survival. Ultimately, patients in our cohort with a 90th percentile ADC value above this threshold were 5.3 times more likely to achieve a CMR.

Importantly, our DWI analysis reaffirms that the relevance of percentile ADC threshold is b value dependent. For our study, 90th percentile ADC threshold from b500 distinguishes a subgroup of patients between the CMR and non-CMR group, while 90th percentile ADC from b1000 obscures such distinction in the high ADC regime. This is because of the noise floor effect of diffusion weighted image signal in the tumor necrosis region [37], which contains almost freely diffusing tissue water (hence high b value diffusion weighting crushes the signal down to noise floor) and contributes most to the 90th percentile ADC. Choosing an optimal b value for similar studies hence depends on ensuring sufficient signal-to-noise ratio for the tumor necrosis region. As it is well-known that ADC value depends on many technical details (e.g., b value) of the diffusion MRI protocol, the threshold found in this study has limited generalizability beyond a single site or across MRI scanner vendors. Further studies drawn from multiple institutions are needed to explore appropriate ADC cutoffs and their role in clinical practice.

Additionally, manual delineation of the tumor region and whole cervix is subject to error, especially at uncertain boundaries. Transformation of regions drawn on T2w anatomical images to ADC maps further compounds any boundary errors. Thus, ADC percentile measures are inevitably sensitive to error in ROI boundary definition and may be influenced by expertise in tumor identification; the impact of this represents another area for future study.

Lastly, the small cohort size of this study limited the strength of some of our statistical measures, hence the exploratory nature of our univariate and multivariate analyses of the potential of DWI in clinical prognostication. Further, while data support that complete metabolic response on PET-CT following therapy is associated with improved overall survival and a reduced risk of isolated local recurrence [10,11,12], we acknowledge that it is remains a surrogate endpoint for actual local failure or disease-free survival. In addition, our diffusion MRI protocols contain imaging parameter variations (e.g., field strength, TR, number of slices, etc.) which is not uncommon in clinical practice (as compared to a prospective research study). Despite this limitation, the quality assurance of our DWI analysis (Fig. 3) rules out common systematic measurement biases in our ADC measurement and adds confidence in designing larger studies in real-world clinical radiology sites.