Ki-67 and the Gleason score (GS) are two of the most important markers used to assess how aggressive a prostate cancer is. Ki-67 is a protein found only in actively dividing cells -- tumors where a high proportion of cells are Ki-67 positive (generally defined as more than 10%) are growing faster and tend to have worse outcomes. The Gleason score describes how abnormal the cancer cells look under a microscope, with scores from 6 to 10 and higher scores indicating more aggressive disease.
Research from the Mayo Clinic has shown that combining Ki-67 expression, Gleason score, and perineural invasion (cancer growing along nerves) provides the most accurate prediction of long-term outcomes in prostate cancer patients. However, all three of these measures currently require invasive biopsy tissue obtained through the rectum -- an uncomfortable procedure that carries risks of infection and bleeding.
This creates a significant clinical need for noninvasive alternatives. MRI-based radiomics -- extracting hundreds of quantitative image features that the human eye cannot see -- offers a potential route to estimating Ki-67 and Gleason score before any tissue is obtained, guiding biopsy decisions and treatment planning.
While previous radiomics studies have investigated either Ki-67 or Gleason score separately, and most used only a single center's data (limiting their generalizability), this study used a three-center multicenter dataset to develop and validate models for both markers simultaneously.
The study enrolled patients from three Chinese hospitals (Gansu Provincial Hospital, the 940 Hospital of Joint Logistics Support Force, and the Second People's Hospital of Gansu Province), providing real-world multicenter variability in patient populations and imaging equipment. After exclusions, 170 patients were included for the Gleason score model and 115 patients for the Ki-67 model (fewer due to some missing immunohistochemistry data).
All patients underwent prostate MRI within 30 days of diagnosis, before any treatment. Three complementary MRI sequences were used: T2-weighted imaging (T2WI) showing anatomy and lesion structure, diffusion-weighted imaging (DWI) reflecting cell density (cancer cells are tightly packed, restricting water movement), and apparent diffusion coefficient (ADC) maps providing a quantitative measure of water diffusion that inversely correlates with tumor density.
Experienced radiologists (5+ years of pelvic MRI experience) and a urologist (30+ years of prostate MRI experience) manually outlined each tumor's boundary on all three sequences, creating three-dimensional regions of interest. When multiple tumors were present, the largest was selected for analysis.
Ki-67 and Gleason score ground-truth labels were established from immunohistochemistry and histological analysis of biopsy or surgical specimens. Ki-67 was classified as high (more than 10% positive cells) versus low, and Gleason score as high (greater than 7) versus low (7 or below).
From each outlined tumor region, 321 radiomics features were mathematically extracted per MRI sequence, giving over 900 features per patient from all three sequences combined. These features fell into three categories: shape features (tumor size, roundness, surface area), first-order features (distribution of pixel brightness values), and texture features (patterns of how pixel values relate to their neighbors, captured by multiple matrix-based calculations including GLCM, GLSZM, and GLRLM).
Because 900+ features far exceed the number of patients available, aggressive feature selection was applied. Highly correlated features were removed using Pearson Correlation Coefficient analysis. Then two different machine learning selection methods were tested: LASSO logistic regression (which mathematically forces most feature weights to zero, retaining only the strongest predictors) and Support Vector Machine (SVM) with recursive feature elimination.
To correct for class imbalance (more high-Ki-67 than low-Ki-67 cases), up-sampling was applied to artificially balance the training set before model building. Without this balancing step, the models showed poor specificity -- they would over-predict the more common class. Balancing significantly improved overall model performance.
The final Ki-67 prediction model used 15 selected features with an SVM classifier, while the Gleason score model used 9 selected features with a LASSO classifier. These models were validated on a held-out test set of patients from the other two hospitals, providing genuine external validation.
The Ki-67 model (SVM classifier) achieved an AUC of 0.884 on the training set and 0.793 on external validation, with sensitivity of 70% and specificity of 85.7%. This means the model correctly identified most patients with high Ki-67 while avoiding many false positives.
The Gleason score model (LASSO classifier) achieved an AUC of 0.827 on training and 0.813 on external validation, with 93.1% sensitivity and 68.2% specificity. The high sensitivity means the model rarely misses high-Gleason tumors, which is the more clinically dangerous error.
Crucially, the radiomics-predicted classifications were not just statistically accurate but clinically meaningful. Kaplan-Meier survival curves showed that patients the radiomics model predicted as having high Ki-67 expression had significantly worse overall survival than those predicted as low -- the same pattern seen with actual lab-measured Ki-67. The same held true for radiomics-predicted Gleason score and survival outcomes (both p less than 0.05).
Decision curve analysis confirmed that both models provided positive net clinical benefit across a range of decision thresholds, meaning they would help clinicians make better decisions than either treating all patients the same or ignoring the imaging information entirely.
The researchers chose different machine learning algorithms for the two prediction tasks, and the choice reflects each algorithm's strengths. SVM was selected for Ki-67 prediction because of its superior performance in image augmentation classification and its ability to handle high-dimensional feature spaces with good generalization -- important when the number of features is large relative to the sample size.
LASSO regression was selected for Gleason score prediction because it simultaneously performs feature selection and regularization, preventing overfitting while identifying the most relevant imaging features. LASSO is particularly effective when predictive features are sparse among a large set of candidates.
The researchers explicitly compared these traditional machine learning approaches against deep learning alternatives (such as VGG and ResNet). They concluded that for their dataset size -- 115-170 patients -- traditional ML models are more appropriate than deep learning, which typically requires thousands of samples for stable training and may overfit on smaller medical imaging datasets.
The fact that texture features from T2WI and DWI sequences dominated the selected feature sets aligns with known biology: prostate tumor heterogeneity -- variation in cell density, gland architecture, and microstructure -- is precisely what these texture measurements capture, and heterogeneity is a known driver of Ki-67 expression and Gleason grade.
The most important finding is that the radiomics-predicted Ki-67 and Gleason score were not just accurate numbers -- they were independently linked to patient survival. Patients predicted by the MRI model to have aggressive disease had significantly worse outcomes than those predicted as low-risk, confirming that the model captures genuine tumor biology rather than just correlating with pathology results by chance.
This opens up the possibility of using MRI radiomics as a tool at multiple points in clinical management: before biopsy to determine biopsy need and targeting, before surgery to plan the extent of resection, and potentially during follow-up to monitor disease behavior without repeat invasive sampling.
The multicenter design strengthens confidence in the model's generalizability compared to single-center studies. Patients from three different hospitals with different patient demographics and MRI scanners were used, and the models maintained good performance across these real-world variations.
Combining radiomics-predicted Ki-67 and Gleason score with other established biomarkers -- as the Mayo Clinic model suggests combining all three with perineural invasion data -- could potentially create even more powerful composite risk tools for personalized prostate cancer management.
MRI-based radiomics models can reliably predict both Ki-67 expression and Gleason score in prostate cancer from noninvasive imaging, with external validation AUCs of 0.793 and 0.813 respectively. The models provide genuine prognostic value, stratifying patients by survival outcomes in a manner consistent with tissue-based pathology.
The multicenter design and the demonstration of survival prognostication are key strengths that distinguish this study from earlier single-center or less validated radiomics work.
The main limitations are the relatively small sample size (115-170 patients), the retrospective design, and the use of manual tumor segmentation which introduces some subjectivity and limits scalability. A prospective multicenter study with a larger cohort is needed before clinical implementation.
Future work should explore automated segmentation, combine radiomics with clinical data in nomogram models, and investigate whether integrating PSMA-PET or genomic data could further boost prediction accuracy. Despite its limitations, this study demonstrates the real potential of MRI radiomics as a practical, noninvasive window into prostate cancer biology.