AI Reading Moves Into Kidney and Prostate Cancer Assessment
The COMFORT study is set to examine whether an artificial intelligence platform can meaningfully support clinicians who assess suspected kidney and prostate tumors. The prospective observational cohort will compare standard medical reading with AI-assisted reading across abdominal CT for renal cancer and multiparametric prostate MRI for prostate cancer. Its scope is broader than simple image classification. Investigators are planning to test whether the platform changes how clinicians interpret lesions, how quickly they read studies, how confidently they use the output, and whether the system behaves consistently across different user groups.
This is not a diagnostic trial in the narrow sense. It is a workflow and decision-support study built around the reality that radiology and urology decisions often depend on structured imaging interpretation, lesion scoring, and repeated judgment under time pressure.
How the Platform Will Be Tested in Real Reading Workflows
The study will enroll an estimated 50 healthcare professionals, specifically residents or specialists in urology or radiology who are proficient in English. Participants will come from multiple sites in Germany, Greece, Italy, and Spain, which makes the project clinically relevant beyond a single center’s reading habits. The cohort design is prospective, and each participant will perform medical reading in two modes: without access to AI output and with access to AI output generated through the COMFORT platform.
In practical terms, the platform is being positioned as a tool for lesion stratification and characterization, not as a replacement for physician judgment. For kidney lesions, the outputs are intended to support assessments such as Bosniak and RENAL scoring. For prostate lesions, the relevant framework includes PI-RADS. Those score systems matter because they are already embedded in everyday practice and shape biopsy decisions, surveillance plans, and downstream referrals. An AI system that can align with those pathways may fit into routine care more naturally than one that produces a generic risk label with little clinical context.
The study protocol also plans separate evaluation for kidney and prostate cancer cases, rather than blending them into a single diagnostic category. That separation is important because the imaging workflow, lesion morphology, and interpretive standards differ between renal CT and prostate mpMRI.
Beyond Accuracy: Fairness, Trust, and Privacy Will Be Measured
COMFORT is structured around several layers of evaluation. Performance on anatomical segmentation, lesion localization, volumetric measurements, classification, and scoring will be compared with clinical reference standards. Agreement between model-generated assessments and reference findings will be assessed across the relevant cancer types. This kind of comparison is essential in imaging AI, where a system can look impressive in one metric and still fail in the real reading room if the outputs are not clinically interpretable.
A separate fairness analysis will examine whether model behavior varies across participant groups. The protocol does not specify the exact groups in the publicly available record, but the intent is explicit: the investigators want to identify differential performance patterns before the platform is treated as broadly reliable. In oncology imaging, unequal behavior across groups can translate into uneven triage, inconsistent confidence, or avoidable variation in lesion characterization.
Privacy is also built into the study question. A privacy causal analysis will examine potential causes of privacy leakage tied to the AI models and characterize their causal effects. That focus reflects a growing concern in medical AI: outputs that look harmless at first glance can still expose sensitive information under certain conditions. In a clinical setting where image-derived outputs may be repeated, stored, or shared across systems, privacy risk is not a side issue. It is part of whether the tool can be responsibly deployed at scale.
Clinician Experience May Decide Whether the Tool Leaves the Pilot Stage
The study will not rely solely on technical metrics. Questionnaires and structured interviews will evaluate acceptability, usability, trustworthiness, perceived information integration, workload management, perceived effectiveness, care management efficiency, and AI literacy from the clinician perspective. That matters because even a strong model can fail in practice if its output is cumbersome, poorly integrated, or too opaque to support decisive action.
A single sentence captures the operational stakes: if clinicians do not trust the output, they will not use it.
The protocol goes further by comparing AI-assisted reading with manual reading in terms of time required and the quality of the resulting assessment. Those comparisons will be examined at individual sites and across the study as a whole, giving the project a real-world implementation dimension. If the platform shortens reading time while preserving or improving interpretive quality, it could support a more efficient diagnostic workflow. If it adds friction, slows reading, or produces output that is technically correct but clinically awkward, its value would be far less certain.
Because the platform is being tested among both urologists and radiologists, the study also has the potential to show whether multidisciplinary use changes the way the same AI output is interpreted. Different specialties often approach the same scan with different thresholds, and an AI system that can remain useful across those perspectives would have a stronger case for clinical adoption.
Why This Study Matters Before Any Results Exist
The estimated study completion date is 2027, and no results are posted yet. Even so, the design is telling. COMFORT is being built around the questions that determine whether imaging AI can be more than a promising software layer: Does it improve decision-making? Does it generalize across settings? Does it behave fairly? Can clinicians actually use it without sacrificing efficiency or trust? And can it do all of that without creating new privacy liabilities?
The inclusion of kidney cancer and prostate cancer in one protocol reflects a common feature of modern imaging AI development: the technical backbone may be shared, but the clinical pathways are not interchangeable. Renal CT interpretation and prostate mpMRI reading each carry distinct scoring systems, thresholds, and downstream consequences. A platform that can support both settings would need to be adaptable enough to respect those differences while still producing consistent, actionable output.
The study will be observed closely because it is not asking whether AI can produce an answer. It is asking whether the answer fits the way clinicians actually work. That is the standard that will matter if this platform is ever to move from a research environment into routine tumor assessment.