Home›Radiology & Imaging› Machine learning predicts Ki-67 in renal tumors with moderate accuracy, AUC 0.85 to 0.86
Machine learning predicts Ki-67 in renal tumors with moderate accuracy, AUC 0.85 to 0.86Machine Learning Models Help Predict Ki-67 in Renal Tumors
Academic radiologyPublished September 6, 2026Study authors: Liu Yu, Zhu Guiqing, Xiu ZhigangPubMed ↗DOI ↗Editorial oversight: Dr. Lars van Dijk, PhD · Surgical, Procedural & Diagnostic
AI-generated summary of the cited source, checked by automated accuracy review.
How we work
Share
Key Takeaway
Consider ML models as complementary, not replacements, for Ki-67 assessment in renal tumors.
This meta-analysis synthesized evidence from 7 retrospective cohort studies to evaluate the diagnostic performance of machine learning (ML) models for predicting Ki-67 expression in renal tumors. The analysis included 1176 patients in training cohorts and 885 in external validation cohorts. The primary outcome was diagnostic performance, measured by sensitivity, specificity, and area under the curve (AUC).
In training cohorts, pooled sensitivity was 0.81, specificity 0.84, and AUC 0.85. In external validation cohorts, sensitivity was 0.83, specificity 0.73, and AUC 0.86. At a 20% pretest probability, the positive post-test probability was 55% and negative post-test probability was 5%, indicating that a negative result is more informative than a positive one.
Subgroup analysis of PyRadiomics-based models (5 studies) showed sensitivity 0.79 and specificity 0.85. Comparing ML algorithms, XGBoost (sensitivity 0.81, specificity 0.84) outperformed Random Forest (0.77/0.83).
The authors noted significant heterogeneity in specificity (I2 > 74%), which limits the precision of pooled estimates. This is a meta-analysis of retrospective cohort studies, not a clinical trial; therefore, the findings represent associations, not causal evidence.
ML models show moderate diagnostic performance and may serve as complementary tools to pathological assessment, but they cannot replace pathological assessment. Clinicians should interpret these results cautiously, as the heterogeneity and retrospective design preclude definitive recommendations for routine clinical use.
How this fits prior evidence
This meta-analysis extends prior coverage on renal tumor management by addressing a diagnostic gap: noninvasive prediction of Ki-67 expression. While earlier items focused on surgical outcomes (e.g., single-port RAPN reducing blood loss, RARN lowering conversion rates, and nephrometry scores showing limited discrimination for adverse outcomes), this review introduces ML-based imaging biomarkers as a potential adjunct. The moderate AUCs (0.85-0.86) are consistent with the limited predictive power seen with nephrometry scores, reinforcing that imaging-based tools are not yet replacements for pathology. The negative post-test probability of 5% at 20% pretest probability suggests ML may help rule out high Ki-67, but heterogeneity (I2 > 74%) tempers confidence.
Researchers conducted a meta-analysis of seven retrospective studies to see if machine learning (ML) could predict Ki-67 expression in patients with renal tumors. Ki-67 is a marker used to help doctors understand how quickly certain cells are dividing. The study looked at a large group of patients across both training and validation cohorts to test the accuracy of these computer models.
The results showed that these models had moderate diagnostic performance. Specifically, the models showed a sensitivity of about 0.81 to 0.83 and a specificity of 0.73 to 0.85. One specific type of model, called XGBoost, performed slightly better than another common model called Random Forest. These results suggest that computer models can be useful for predicting markers that are usually only found through tissue samples.
It is important to note that this was a meta-analysis of older data, not a new clinical trial. The study also noted significant differences in how well the models performed in different settings. While these tools could eventually help doctors, they are not intended to replace the standard pathological exams currently used to diagnose and manage renal tumors.
What this means for you:
Machine learning models show moderate accuracy in predicting Ki-67 levels in renal tumors as a possible tool.
Common questions
What is Ki-67 and why is it measured in renal tumors?
Ki-67 is a marker used to help doctors understand how quickly cells are dividing. In the context of renal tumors, it helps doctors assess the behavior of the tumor. This study looked at using machine learning to predict these levels noninvasively.
How accurate were the machine learning models?
The models showed moderate diagnostic performance. In the validation group, the models had a sensitivity of 0.83 and a specificity of 0.73. One specific model, XGBoost, showed slightly higher performance than the Random Forest model.
Can these computer models replace a biopsy?
No, these machine learning models cannot replace pathological assessment. The study suggests they may serve as complementary tools to help doctors, but they are not a replacement for standard medical testing.
RATIONALE AND OBJECTIVES: To systematically evaluate the diagnostic performance of machine learning (ML) models for predicting Ki-67 expression in renal tumors and assess their potential for clinical translation.
MATERIALS AND METHODS: We systematically searched PubMed, Web of Science, Cochrane Library, and Embase up to November 2025. Studies using ML to predict Ki-67 with immunohistochemistry as reference standard were included. Quality was assessed with QUADAS-2. A bivariate random-effects model pooled diagnostic metrics. Meta-regression and subgroup analyses explored heterogeneity.
RESULTS: Seven retrospective cohort studies were included, comprising 1176 patients in the training cohorts and 885 patients in the external validation cohorts. Pooled sensitivity, specificity, and area under the curve (AUC) were 0.81, 0.84, and 0.85 for training cohorts, and 0.83, 0.73, and 0.86 for validation cohorts. At 20% pretest probability, positive/negative predictions modified post-test probabilities to 55%/5%. Specificity showed substantial heterogeneity (I² > 74%). Meta-regression identified tumor type, Ki-67 cut-off, feature extraction software, and ML algorithm as heterogeneity sources. Subgroup analyses showed: RCC-specific studies and 5% Ki-67 cut-offs yielded higher estimates; PyRadiomics was most used (5 studies; sensitivity 0.79, specificity 0.85); eXtreme Gradient Boosting (XGBoost) showed numerically higher sensitivity and specificity than Random Forest (0.81/0.84 vs. 0.77/0.83).
CONCLUSION: ML demonstrates moderate diagnostic performance for noninvasive Ki-67 prediction in renal tumors. Diagnostic accuracy is influenced by tumor type, Ki-67 cut-off, feature extraction software, and algorithm selection. While current models cannot replace pathological assessment, they may serve as complementary tools. Future research should prioritize algorithm optimization, technical standardization, and prospective multi-center validation to enhance clinical applicability.
