Summary
A medRxiv preprint evaluates BanffNET indices that summarise kidney-transplant biopsy injury and distinguish several Banff pathology categories. The model was assessed across 6,407 biopsies and a multi-reader study involving 67 pathologists.
A computational pathology model called BanffNET has been used to convert automated lesion scores from kidney-transplant biopsies into four continuous indices describing different forms and stages of graft injury, according to a medRxiv preprint.
The indices were evaluated using 6,407 biopsy records, a multi-reader study involving 67 pathologists, molecular biopsy signatures and time to kidney graft failure. The researchers report that three of the indices distinguished important Banff pathology categories in validation data, with area-under-the-curve (AUC) values ranging from 0.80 to 0.92.
The work is a preprint describing development and evaluation of a diagnostic prediction model. The authors propose that the indices could complement, rather than replace, pathologists’ assessment of transplant biopsies.
From categorical diagnoses to continuous injury measures
Kidney-transplant biopsies are commonly interpreted using the Banff classification, which organises microscopic findings into lesion scores and diagnostic categories. The preprint identifies substantial inter-observer variability and potential misinterpretation as challenges for this approach.
BanffNET was used to generate 17 automated lesion scores from whole-slide biopsy images. The researchers then applied a penalised regression method to condense those scores into four measures:
- a TCMR/TI Index, relating to T-cell-mediated rejection and inflammation-related injury;
- an AMR/MVI Index, relating to antibody-mediated rejection and microvascular inflammation;
- an Activity Index, representing active injury features; and
- a Chronicity Index, representing longer-term structural damage.
Unlike a single categorical diagnosis, a continuous index can express the degree to which a biopsy displays a particular injury pattern. That may provide a more detailed description of tissue changes alongside the pathologist’s diagnosis.
How the model was evaluated
The development cohort contained 2,544 biopsies, while a separate validation cohort contained 3,863. The researchers compared BanffNET indices with Banff diagnoses and histological indices in both cohorts.
They also used a multi-reader cohort of 36 biopsies scored by 67 pathologists. Because individual pathologist assessments can vary, the study examined how closely the indices related to multi-observer reference standards. Additional comparisons involved molecular signatures derived from validation-cohort biopsies and time to kidney graft failure in the training and validation cohorts.
For the main diagnostic comparisons, the TCMR/TI Index, AMR/MVI Index and Activity Index showed strong discrimination of Banff TCMR, Banff microvascular inflammation and any Banff diagnosis, respectively. Their validation AUCs ranged from 0.80 to 0.92; an AUC of 1 represents perfect separation between groups, while 0.5 represents performance equivalent to chance.
The authors report that, in most comparisons, the BanffNET indices explained as much or more variability in multi-observer diagnoses, graft-failure timing and molecular biopsy diagnostics as pathologist-assigned diagnoses and histological indices.
Potential role in transplant pathology
The reported results suggest that automated continuous measures may capture both the type and severity of injury in a biopsy. This could be useful when a specimen contains overlapping features or when a categorical diagnosis does not fully represent the range of tissue abnormalities.
The indices are presented as reproducible descriptors of whole-slide images that could augment routine pathology review. Their associations with graft-failure timing and molecular signatures also connect the image-derived measures with clinically and biologically relevant characteristics of transplant injury.
The evidence is currently from a medRxiv preprint and concerns model development, diagnostic discrimination and associations with pathology and graft outcomes. It does not evaluate a prospective clinical workflow in which treatment decisions are made using the indices. The practical value of the system will therefore depend on how it performs across additional centres and biopsy types, and how transplant teams incorporate its continuous measurements into pathologist-led assessment.