Ki67 is a protein found inside cells that is active whenever a cell is dividing. Its levels serve as a direct readout of how fast a tumor is growing -- tumors with high Ki67 expression are more aggressive, more likely to spread, and associated with worse outcomes.
In prostate cancer specifically, patients with high Ki67 expression (defined as 10% or more of cancer cells staining positive) show poorer tumor differentiation and a 2.62-fold higher risk of biochemical recurrence compared to those with low Ki67. It is also linked to distant metastasis and shorter disease-free survival.
Currently, Ki67 levels can only be measured through biopsy tissue samples analyzed in a laboratory -- a process that is invasive, prone to sampling errors, time-consuming, and subject to variation between different pathologists and labs. This creates a need for noninvasive alternatives.
Multiparametric MRI, which combines several imaging sequences to give detailed information about prostate tissue, is already a standard tool for prostate cancer diagnosis. This study asks whether deep learning analysis of these MRI images can predict Ki67 expression before any surgery or biopsy takes place.
The study included 229 patients with pathologically confirmed prostate cancer from two medical centers in China, collected between 2019 and 2021. All patients underwent multiparametric MRI prior to biopsy or surgery, providing three complementary image types: T2-weighted imaging (T2WI), diffusion-weighted imaging (DWI), and contrast-enhanced T1-weighted imaging (CE-T1).
A radiologist with 10 years of experience manually selected the region of interest (the largest cross-section of each tumor) from each MRI sequence, and a pathologist with 15 years of experience confirmed these selections against biopsy results. Cases were then split: 70% for model training, 30% for internal testing, and the full dataset from the second hospital served as an external validation set.
Ki67 expression was measured from surgical or biopsy specimens using standard laboratory staining. Patients were classified into two groups: low Ki67 expression (less than 10%) and high Ki67 expression (10% or greater). Approximately 25-38% of patients in each data split had high Ki67 expression.
Clinical features that differed significantly between the high and low Ki67 groups in the training set included MRI tumor staging (mrT), lymph node involvement, distant metastasis, total PSA levels, tumor size, location within the prostate, and ADC values -- a measure of water movement in tissue that reflects cellular density.
The researchers tested three distinct deep learning neural network designs, each with different strengths. ResNet101 is a 101-layer residual network that uses shortcut connections to prevent information loss in very deep networks, enabling it to capture complex image patterns. InceptionV3 uses parallel convolution operations of different sizes, efficiently handling features at multiple scales simultaneously. DenseNet121 connects each layer to every other layer in its dense blocks, allowing maximum reuse of learned features with fewer parameters.
Each network automatically extracted thousands of deep imaging features from the three MRI sequences: ResNet101 and InceptionV3 each produced 6,144 features per patient, while DenseNet121 produced 3,000. These raw features capture visual patterns in the images that are too subtle or complex for human radiologists to detect by eye.
To prevent overfitting -- where a model learns noise rather than true patterns -- two sequential feature selection steps were applied. First, a method called mRMR identified the most informative features while removing redundant ones. Then, LASSO regression with 10-fold cross-validation further reduced the feature set to only those most strongly associated with Ki67 expression. The final selected feature counts were 43, 39, and 78 for the three networks respectively.
Each deep learning model produced a single numerical score (called a deep learning radiomic signature, or DLRS) for each patient that summarized the imaging evidence for high versus low Ki67 expression. These scores were significantly different between the high and low Ki67 groups across all data sets.
In parallel with the deep learning models, the researchers built a traditional clinical prediction model using only standard clinical and imaging measurements. Logistic regression analysis tested 14 candidate variables to identify which independently predicted Ki67 expression.
Only two variables emerged as independent predictors: MRI tumor stage (mrT) and the ADC value of the tumor. Both make biological sense -- higher tumor stage reflects more aggressive local growth, and lower ADC values indicate denser cellular packing, both characteristics of rapidly proliferating tumors.
To maximize predictive power, each deep learning score was then combined with these two clinical variables to create three nomograms -- visual scoring tools (Nomogram-Resnet, Nomogram-Inception, Nomogram-Densenet) that translate combined clinical and imaging inputs into an estimated probability of high Ki67 expression.
Nomograms allow a physician to visually read off a predicted probability for an individual patient by finding each variable's value on a scale and summing the points. This makes the combined model more accessible and interpretable in clinical practice compared to a pure computational output.
The clinical model using only standard measurements achieved AUC values of 0.79, 0.71, and 0.75 on the training, internal validation, and external validation sets respectively. While reasonable, these values indicate meaningful diagnostic gaps.
All three deep learning models dramatically outperformed the clinical model. AUC values for the DLRS models ranged from 0.939 to 0.983 across datasets, representing near-perfect discrimination between high and low Ki67 tumors. The improvement over the clinical model was statistically significant for all three deep learning architectures (p less than 0.01).
The combined nomograms performed best of all, achieving AUCs from 0.952 to 0.993. The Nomogram-Resnet, for example, reached an AUC of 0.993 on external validation -- meaning it correctly ranked almost every patient. Calibration curves confirmed that the models' predicted probabilities closely matched actual observed Ki67 expression rates.
Decision curve analysis further confirmed that the deep learning models and nomograms provided greater net clinical benefit than the clinical model across a wide range of decision thresholds. Notably, results were consistent across datasets from two hospitals using different MRI scanners, demonstrating strong generalizability.
Previous studies attempted to predict Ki67 expression using either quantitative MRI parameters or handcrafted radiomic features -- manual measurements from images that require expert selection. These approaches required extensive manual work for each patient and were not externally validated, leaving their reliability uncertain.
The key advantage of the deep learning approach in this study is automation: the networks automatically learn which image patterns predict Ki67 expression, eliminating subjective feature selection and reducing radiologist workload. The simplification to single-layer segmentation (rather than manual outlining of the entire tumor slice by slice) also makes the process faster and more practical.
Three different architectures were tested rather than just one, providing confidence that the results reflect genuine predictive power rather than artifacts of one particular model design. All three performed similarly well, suggesting the underlying image patterns predicting Ki67 are robust and not specific to one neural network approach.
The finding that scanner differences between the two hospitals did not significantly impair performance is particularly important for future clinical use, as it suggests the models can generalize across different imaging equipment -- a common challenge in radiology AI research.
The most significant clinical implication of this work is the possibility of obtaining detailed prognostic information before surgery or biopsy. Knowing whether a patient's tumor has high Ki67 expression -- indicating aggressive, rapidly dividing cancer -- could directly influence treatment decisions.
Patients identified as likely having high Ki67 tumors might be prioritized for more aggressive treatment regimens, closer surveillance, or earlier initiation of systemic therapy. Those with low Ki67 scores might be candidates for active surveillance or less aggressive intervention, potentially avoiding overtreatment.
The nomogram format makes these AI-derived insights actionable: by summing scores from MRI staging, ADC value, and the deep learning score, a physician can estimate a specific patient's probability of high Ki67 expression before tissue is obtained. This represents a new category of imaging-based biomarkers that complement but do not replace pathology.
Future clinical integration would require prospective validation in larger, more diverse patient populations and eventual incorporation into standard radiology reporting workflows. The study authors note that automatic segmentation technology -- removing the remaining manual step -- should be explored in follow-up work.
This study successfully developed three deep learning models and three combined nomograms for noninvasive preoperative prediction of Ki67 expression in prostate cancer. All models demonstrated strong performance with robust internal and external validation, suggesting genuine clinical potential.
The approach offers a practical workflow: radiologists already performing routine multiparametric MRI would obtain Ki67 predictions automatically without additional patient burden, additional imaging time, or the inherent sampling limitations of tissue biopsy.
The main limitations acknowledged by the authors include a relatively small sample size with more high-Ki67 than low-Ki67 patients, which may bias the models and requires validation in larger balanced cohorts. Manual segmentation of tumor regions also introduces some subjectivity, and future work should address this with automated segmentation.
Overall, this research contributes a validated, easy-to-use deep learning tool that could meaningfully advance personalized preoperative planning for prostate cancer patients by making molecular prognostic information accessible from standard imaging.