The clinical motivation
A meaningful number of kidney lesions turn out to be benign or indolent, and a meaningful number of biopsies and surgeries are therefore avoidable in hindsight. A non-invasive test that distinguishes tumour biology before the operation would change that calculus, both by reducing unnecessary procedures and by improving surgical planning through pre-operative tissue assessment.
Intravoxel incoherent motion (IVIM) is a candidate. The model separates the diffusion-weighted signal into tissue diffusivity Dt, which should reflect cellular density and organisation, and two microvascular terms: perfusion fraction fp and pseudo-diffusivity Dp. IVIM has shown promise for renal tumour diagnosis and subtyping, but validation against detailed histopathology has been thinner than the clinical claims resting on it.
The gap this study addresses is specific. Prior work largely compared IVIM against diagnosis or against other imaging. Here the comparison is against the resected tissue, scored by pathologists and quantified by multiplex immunofluorescence, in the same patients.
Design
Renal mass patients consented to pre-operative IVIM-DWI and then underwent partial nephrectomy, so both imaging and tissue were available per patient. Twenty-seven patients completed IVIM analysis, twenty-four had qualitative histopathology and twenty-two had quantitative measurements.
Imaging
Abdominal imaging on a 3T MAGNETOM Prisma, respiratory-gated single-shot EPI, 2.13 × 2.13 × 5 mm resolution, eleven b-values from 0 to 800 s/mm² in three directions, six-minute acquisition. Tumours were manually segmented in FireVoxel and a segmented IVIM model was fitted for Dt, fp and Dp.
Rather than reducing each tumour to three mean values, histogram analysis extracted mean, variance, skewness, kurtosis, entropy and nine percentiles for every parameter. This is the part that matters methodologically. A tumour is heterogeneous, and the mean discards exactly the heterogeneity that distinguishes one tumour from another.
Histology
Two independent readouts from the same tissue. Pathologists scored vascularity (1 to 3), inflammation (0 to 1), cyst presence (0 to 3) and fibrosis (0 to 3) on H&E sections. Separately, multiplex immunofluorescence was quantified in HALO across five channels: DAPI for cell nuclei, CD31 for mature vessels, CD34 for immature vessels, pan-cytokeratin for epithelial cells, and unstained "negative" area. Measurements covered stain areal fractions and microvasculature descriptors.
Association was tested with Spearman rank correlation at p ≤ 0.05, chosen over Pearson because several histology readouts are ordinal scores rather than continuous quantities and the relationships were not assumed linear.
Results
Against pathologist scoring
| Histology score | Correlated IVIM parameter | Reading |
|---|---|---|
| Vascularity | Dp mean, Dp entropy | Sensitivity to micro-vascularity |
| Inflammation | Dt mean, (Dp×fp) variance | Restricted diffusion with heterogeneous perfusion |
| Cyst presence | Dt mean, Dt variance | Elevated diffusivity in fluid-filled space |
| Fibrosis | fp mean, Dt variance | Indirect, see below |
The fibrosis result is the one that needs care. A correlation with perfusion fraction is not obviously mechanistic, and the most plausible reading is indirect: fibrotic tissue relaxes fast enough to drop out of the measured signal, which shifts the apparent weighting toward vascular flow volumes. That is an interpretation, not a demonstration, and it is worth flagging as such.
Against quantitative immunofluorescence
Dt mean correlated negatively with DAPI areal fraction and positively with unstained areal fraction. Both point the same way: water diffuses less freely where nuclei are densely packed, and more freely in minimally restricted space. Dp at the 5th percentile correlated positively with CD34 areal percentage, tying pseudo-diffusivity to immature vessel content, with p < 0.01 throughout.
Dt mean, the most conventional IVIM parameter, proved sensitive to most quantitative histologic features including micro-vascularisation. That breadth is a mixed result: sensitive but not specific, which limits how much diagnostic weight a single mean value can carry on its own. The radiomics were more discriminating, and Dt radiomics in particular were the most sensitive to microstructural features.
What this supports, and what it doesn't
- Supports: IVIM parameters carry real information about renal tumour microstructure, validated against tissue rather than against another imaging method. That is the evidence a non-invasive biomarker needs before it can carry clinical weight.
- Supports: distributional features beat means. Entropy, variance and percentiles surfaced relationships that mean values alone did not, which argues for treating parameter maps as distributions rather than collapsing them to a number.
- Does not support: voxel-level inference. Imaging and tissue were never spatially registered, and the study establishes agreement between tumour-level summaries only. Nothing here says what any individual voxel contains.
- Caveat: the cohort is small, single-institution and single organ, and microvascular features were relatively sparse on histology, which may have weakened their measured relationships with global MR features. These are grounds for a larger study, not a substitute for one.
Presented as an oral at ISMRM 2026, in the Advances in Renal MRI session. Work carried out at the Center for Advanced Imaging Innovation and Research (CAI2R), NYU Grossman School of Medicine, with Dr. Eric Sigmund, Dr. Nima Gilani and Dr. Hersh Chandarana. Supported by NIH R01CA245671 and NIBIB P41 EB017183. A companion study on individual kidney function decline after partial nephrectomy was also presented at ISMRM 2026.