The distinction between the two families of imaging measurements used in this study is worth unpacking, because they interrogate fundamentally different aspects of tumour biology. Dynamic contrast-enhanced MRI tracks the behaviour of a gadolinium-based contrast agent as it moves from the blood compartment into the tumour’s extracellular space and back again. When fitted to a pharmacokinetic model — in this case the extended Tofts model, which has been the workhorse of quantitative perfusion analysis for more than two decades — the resulting time–concentration curves can be decomposed into parameters with direct physiological meaning. Ktrans reflects the volume transfer constant, a composite measure of blood flow, capillary permeability and capillary surface area per unit volume of tissue. Kep is the rate constant describing washout of contrast from the extracellular space back into the vasculature, and Ve represents the fractional volume of the extravascular extracellular space per unit volume of tissue. Diffusion-weighted imaging, by contrast, probes the random Brownian motion of water molecules within tissue. The apparent diffusion coefficient quantifies the degree to which that motion is restricted, which in tumours is largely a function of cellularity: densely packed cells with intact membranes restrict water movement and lower the ADC, while necrotic or oedematous regions permit freer diffusion and raise it.

That the two modalities should carry complementary information is biologically plausible. Aggressive tumours tend to be both highly vascularised — with leaky, disorganised neovasculature that elevates Ktrans — and densely cellular, which depresses ADC. The negative correlations observed between ER and PR status and both Ktrans and ADC in this cohort fit that expectation: hormone receptor-positive tumours are, on average, more differentiated and less angiogenically active than their receptor-negative counterparts. The positive correlation between Ki-67 and both Ktrans and Kep is similarly coherent, since a high proliferation index drives both metabolic demand and angiogenic signalling, and rapidly proliferating tissue typically shows fast contrast washout. The HER-2 findings are more nuanced. The positive association with Ve and ADC alongside a negative association with Kep suggests that HER-2-amplified tumours may have a larger and more freely diffusing extracellular compartment, a pattern that has been reported in some earlier single-centre series but remains inconsistent across the literature.

The statistical rigour applied to these correlations deserves comment, because it is unusually careful for this genre of study. With sixteen separate Spearman tests conducted across four biomarkers and four imaging parameters, the probability of observing at least one nominally significant result by chance alone is substantial. By applying a Bonferroni correction that sets the per-test significance threshold at approximately 0.003, the authors guard against the common pitfall of over-interpreting marginal associations. That all the reported correlations survived this correction, with P values below 0.001, lends credibility to the finding that these are genuine population-level relationships rather than statistical artefacts — even if the effect sizes, as the authors themselves note, are modest.

The reproducibility analysis is another methodological strength that merits attention. Quantitative imaging biomarkers are only useful clinically if they can be measured consistently, both by different observers and by the same observer at different times. The intraclass correlation coefficients reported here, all at or above 0.85 across 100 independently re-measured lesions, indicate excellent agreement. This is not a trivial achievement: region-of-interest placement in breast MRI is notoriously sensitive to reader technique, lesion heterogeneity, and the choice of which slices and which enhancement phases to include. The fact that the measurements were performed on a vendor-neutral post-processing platform is also relevant, since it suggests the approach could in principle be replicated at other centres without being tied to a specific scanner manufacturer’s proprietary software.

The TNBC classification results are arguably the most clinically consequential finding. Triple-negative breast cancer, defined by the absence of oestrogen receptor, progesterone receptor and HER-2 amplification, accounts for roughly 15 to 20 percent of breast cancers and carries a disproportionately poor prognosis relative to its incidence. It is the subtype for which no targeted therapy exists, for which neoadjuvant chemotherapy remains the mainstay of treatment, and for which preoperative identification has the greatest potential to alter management. A single-parameter AUC of 0.826 for Kep is respectable but not decisive; what is striking is the performance of the multivariable model. By combining Ktrans, Kep and an inverted Ve — inverted because Ve discriminates in the opposite direction, with TNBC showing the lowest values — the model achieves an AUC of 0.891, a statistically significant improvement over any single parameter alone as confirmed by the DeLong test.

The inversion of Ve is a small but instructive detail. In many radiomics and machine-learning applications, variables that discriminate in the opposite direction from the majority are simply left out or, worse, their contribution is misinterpreted. Here the authors recognised that a low Ve — a contracted extracellular space — is itself a signature of densely packed, poorly differentiated tissue, and that this information is complementary rather than redundant with the perfusion parameters. Including it in inverted form allows the logistic regression to exploit the full directional structure of the data rather than discarding a genuinely informative signal.

It is important to situate these findings within the broader landscape of breast imaging research. Numerous prior studies have attempted to link DCE-MRI parameters or ADC values to molecular subtype, with results that vary considerably across institutions, scanner field strengths, contrast protocols and pharmacokinetic models. Some groups have reported that Luminal A tumours show lower Ktrans than other subtypes; others have found no significant difference. ADC-based discrimination has been similarly inconsistent, with some meta-analyses suggesting modest but significant differences between TNBC and non-TNBC tumours and others finding the effect too small to be clinically useful. The heterogeneity of prior work is precisely what the authors of the present study identify as a limitation of the existing literature, and their decision to analyse a single large cohort with a uniform 3.0T protocol and formal multivariable modelling is a deliberate attempt to address that gap.

The cohort size of 746 patients is substantial by the standards of this literature, and the subtype distribution — 149 Luminal A, 400 Luminal B, 88 HER-2 overexpression and 109 TNBC — is broadly representative of what would be expected in an unselected Chinese breast cancer population, with Luminal B tumours predominating as they do in most large series. The imbalance in subtype numbers, however, means that the model’s performance for the rarer subtypes, particularly HER-2 overexpression, is less well characterised than for TNBC. The authors wisely confined their multivariable modelling to TNBC rather than attempting a four-way classification, which would have required substantially more events per variable to remain statistically robust.

The retrospective design introduces its own set of caveats. All patients underwent imaging at a single institution over a three-year period, and the imaging protocol, while standardised within that institution, may not generalise to centres using different field strengths, contrast agents, injection rates or pharmacokinetic modelling software. The authors acknowledge this by framing their conclusions as hypothesis-generating and calling for prospective external validation. That framing is appropriate and, frankly, more honest than much of the literature in this space, where the gap between reported AUCs and clinical utility is often glossed over.

What would external validation need to demonstrate? At minimum, the model’s AUC would need to be reproduced in an independent cohort, ideally from a different institution and ideally with a different scanner vendor, to confirm that the parameter values and their relationships to molecular subtype are not artefacts of local protocol. Beyond that, the model would need to demonstrate clinical utility — that is, that using it to preoperatively flag suspected TNBC actually changes management decisions in a way that improves outcomes. This is a much higher bar than statistical discrimination, and it is the bar that ultimately determines whether a biomarker transitions from research tool to clinical practice.

The relationship between imaging biomarkers and tumour biology is also worth considering from a mechanistic standpoint. Ktrans, as a composite of flow, permeability and surface area, is influenced by factors beyond tumour aggressiveness alone. Interstitial pressure within the tumour, the degree of stromal fibrosis, the presence of necrosis and even systemic factors affecting cardiac output and blood pressure can all modulate contrast kinetics. This means that while Ktrans correlates with biological features such as grade and proliferation, it is not a direct readout of any single molecular process. Similarly, ADC is affected not only by cellularity but by the integrity of cell membranes, the viscosity of the extracellular fluid and the presence of macromolecules. The weak-to-moderate correlations observed here are therefore not surprising; they reflect the reality that imaging parameters are indirect, multi-factorial proxies for underlying biology rather than direct molecular measurements.

The choice of the extended Tofts model over alternatives such as the shutter-speed model or reference-tissue methods is standard but not without consequence. The extended Tofts model assumes that the contrast agent exchanges between vascular and extravascular spaces rapidly relative to the temporal resolution of the acquisition, an assumption that may not hold for all tumour types and that can bias parameter estimates in tissues with very fast or very slow kinetics. The arterial input function, which is required to convert signal intensity changes into concentration values, is another source of variability. Whether it was measured directly from a vessel or derived from a population-averaged template affects the absolute values of Ktrans and Kep, even if the relative comparisons between subtypes remain valid.

The supplementary materials indicate that the authors conducted a Pearson sensitivity analysis alongside their primary Spearman correlations, which is a sensible robustness check given that Spearman makes no distributional assumptions but sacrifices some statistical power when the underlying relationship is genuinely linear. The Shapiro–Wilk and Levene’s tests reported in Table S1 confirm that the distributions of the imaging parameters were not normally distributed, justifying the use of non-parametric methods as the primary analysis. The post-hoc power analyses in Table S5 address the perennial concern about whether non-significant findings reflect true absence of association or simply insufficient sample size.

Looking forward, the convergence of quantitative MRI with machine learning approaches — hinted at in the article’s suggested related subjects — represents a natural next step. The multivariable logistic regression used here is a simple, interpretable model, which is appropriate for a first demonstration but may underperform relative to more flexible approaches that can capture non-linear interactions between parameters. Radiomics pipelines that extract hundreds of texture and shape features from the same images, combined with deep learning architectures that can learn representations directly from pixel data, have shown promise in related tasks such as predicting lymph node status and response to neoadjuvant chemotherapy. Whether such approaches can meaningfully improve upon the 0.891 AUC reported here, or whether they simply add complexity without adding discriminative power, remains an open question that future studies will need to address with the same methodological care demonstrated in this work.

For clinicians, the practical takeaway is that quantitative DCE-MRI parameters, measured reproducibly at 3.0T, carry real but limited information about breast cancer biology at the population level. They are not a substitute for biopsy and immunohistochemistry, and they are unlikely to become one in the foreseeable future. What they may offer, once validated prospectively, is a non-invasive adjunct — a way to raise or lower the pre-test probability of TNBC before the pathology report arrives, to inform the urgency of genetic counselling referrals, or to help stratify patients in clinical trial settings where molecular subtype information is needed before tissue is available. The gap between a well-performing statistical model and a clinically actionable tool is wide, but studies of this calibre, with their attention to reproducibility, statistical correction and honest framing of limitations, are the necessary first steps across it.

Subject of Research: Correlation between quantitative parameters and apparent diffusion coefficient of 3.0T dynamic contrast-enhanced MRI and prognostic factors and molecular classification of breast cancer

Article Title: Correlation between quantitative parameters and apparent diffusion coefficient of 3.0T dynamic contrast-enhanced MRI and prognostic factors and molecular classification of breast cancer

Article References: Fu, J., Jiang, Q., Sun, C., & Pan, J. (2026). Correlation between quantitative parameters and apparent diffusion coefficient of 3.0T dynamic contrast-enhanced MRI and prognostic factors and molecular classification of breast cancer. BMC Cancer. https://doi.org/10.1186/s12885-026-16485-2

Image Credits: AI Generated

DOI: 10.1186/s12885-026-16485-2

Keywords: Correlation, between, quantitative, parameters, apparent, diffusion, coefficient, dynamic, contrast-enhanced, prognostic, factors, molecular