Prostate cancer is the most common cancer in men in Western countries. While many cases are slow-growing and low-risk, others require immediate treatment. The clinical challenge is accurately distinguishing aggressive tumors from indolent ones that can be safely monitored.
The current standard for prostate cancer detection involves systematic biopsy, a procedure that samples multiple regions of the prostate at random. This approach is invasive, prone to infection, and still misses approximately one in five clinically significant cancers while over-detecting low-risk tumors that will never cause harm.
Multiparametric MRI (mpMRI) has transformed prostate cancer evaluation by combining structural T2-weighted imaging with functional sequences including diffusion-weighted imaging (DWI) and dynamic contrast enhancement (DCE). This combination reveals tumor location, extent, and aggressiveness in a non-invasive way.
Despite mpMRI's advantages, interpretation relies heavily on radiologist expertise and produces variable results between readers and institutions. An automated, objective system that standardizes interpretation and provides reproducible aggressiveness estimates could substantially improve how prostate cancer is managed.
Researchers from the Candiolo Cancer Institute and San Luigi Hospital in Italy developed a fully automatic computer-aided detection (CAD) system that analyzes mpMRI to detect prostate cancer lesions and classify their aggressiveness without any manual input from a radiologist.
The system operates in two stages: first, it automatically segments both the prostate gland and candidate tumor lesions; second, it extracts radiomics features from the segmented regions and classifies each lesion as either low-aggressiveness (Gleason Grade 2 or below) or high-aggressiveness (Gleason Grade 3 or above).
A key design goal was multi-center and multi-scanner validation. The system was trained on data from one institution and one scanner, then tested on data from a second institution using different scanner equipment, specifically to evaluate whether the approach generalizes beyond the environment where it was built.
The clinical purpose was to provide oncologists with an objective, reproducible risk score for each lesion that could reduce unnecessary biopsies for indolent tumors while ensuring aggressive cancers are never missed or misclassified as benign.
The study enrolled 131 patients with biopsy-proven prostate cancer who all underwent radical prostatectomy within three months of their MRI scan. Post-surgical pathology provided the ground truth for each tumor's Gleason Grade, allowing precise comparison of the AI's predictions against the actual diagnosis.
All patients had complete mpMRI including T2-weighted imaging, diffusion-weighted imaging (DWI) with corresponding ADC maps, and dynamic contrast-enhanced (DCE) imaging. Across all patients, 149 clinically significant tumors were identified and mapped.
Patients from Candiolo Cancer Institute (center A) were used for training and internal validation, while patients from San Luigi Hospital (center B) provided an independent external validation set. The two centers used different scanner models and acquisition protocols, making this a genuinely challenging cross-site test.
Tumor regions were manually delineated on the MRI by expert radiologists using pathological maps from the surgical specimens as reference. These expert-drawn contours served as the ground truth for training the detection and segmentation modules of the system.
Radiomics is the process of extracting quantitative numerical features from medical images that describe texture, shape, and intensity patterns invisible to the naked eye. From each segmented tumor region, the system extracted 92 radiomics features from the ADC maps and T2-weighted images.
The features fell into three categories: volumetric and shape features (lesion volume, surface area, compactness), intensity statistics (mean, variance, skewness of pixel values), and texture features derived from Gray-Level Co-occurrence Matrices (GLCM) and Gray-Level Run-Length Matrices (GLRLM), which capture how pixel intensities are spatially organized within the tumor.
With 92 candidate features but limited training data, overfitting is a serious risk. The researchers applied statistical feature selection, reducing the set to the 6 most informative features: lesion volume, two T2-weighted texture features (difference average GLCM and a GLRLM feature), and three ADC-derived texture features (JointMax GLCM and two GLRLM features).
These 6 features were then used to train a Support Vector Machine (SVM) classifier with a third-order polynomial kernel, a well-established machine learning approach that finds the optimal mathematical boundary separating low-aggressiveness from high-aggressiveness tumor classes.
For distinguishing low-aggressiveness (Gleason Grade 2 or below) from high-aggressiveness (Gleason Grade 3 or above) tumors, the system achieved an AUC of 0.96 in training and 0.81 in external validation. Both values represent clinically meaningful discriminative ability, especially given the challenge of generalizing to a different institution and scanner.
Rather than providing a simple binary classification, the system assigned each lesion to one of three risk tiers: indolent (score falling in the lowest 15% below the decision cutoff), indeterminate (within 15% of the cutoff boundary), and aggressive (above the cutoff). This three-tier approach acknowledges clinical uncertainty and avoids overconfident misclassification.
The most critical safety result was that no aggressive tumor was classified as indolent in either the training or external validation sets. All false negatives, meaning aggressive tumors the system did not correctly categorize as aggressive, fell into the intermediate indeterminate category rather than being incorrectly reassigned to the safe indolent tier.
The system provided definitive aggressiveness classification for 51 to 53% of lesions (either clearly indolent or clearly aggressive), with the remaining lesions falling into the indeterminate zone and flagged for closer evaluation. This approach prioritizes patient safety over classification rate.
The selection of lesion volume as one of the six key features aligns with the known clinical correlation between tumor size and aggressiveness: larger tumors are generally more aggressive. However, volume alone is insufficient, which is why the additional texture features provide independent information.
The selection of features from both T2-weighted images and ADC maps is biologically meaningful. T2 signal reflects tissue water content and cellular architecture, while ADC reflects water diffusion restriction, which correlates with cellular density. High-grade tumors have denser cell packing, restricting diffusion and producing lower ADC values and more homogeneous texture patterns.
The GLRLM features appearing in both T2 and ADC channels capture the spatial runs of similar pixel values within the tumor, describing how homogeneous or heterogeneous the texture is across the lesion. Aggressive tumors tend toward more uniform, dense textures that are captured by these run-length statistics.
The fact that only 6 features provided near-optimal classification suggests that prostate tumor aggressiveness leaves a consistent radiomics signature across different patient populations and scanner types, validating the biological basis for the quantitative imaging approach.
The three-tier classification system was specifically designed with clinical safety as the primary constraint. In cancer care, a false negative (calling an aggressive cancer safe) is far more harmful than a false positive (calling a benign lesion suspicious). The indeterminate category serves as a buffer against false negatives.
By ensuring all misclassified aggressive tumors fell into the indeterminate category rather than the indolent one, the system preserves the physician's ability to catch every aggressive cancer. The indeterminate lesions are then flagged for additional review, targeted biopsy, or closer surveillance, rather than being dismissed.
The 51 to 53% definitive classification rate may seem low but represents a meaningful clinical gain: for roughly half of all lesions, the system provides high-confidence guidance. For patients with tumors in the definitive indolent category, unnecessary biopsy could potentially be avoided, reducing procedural risk and patient anxiety.
Compared to prior CAD systems for prostate cancer, this study provides a more rigorous validation by testing across two institutions with different scanners, demonstrating that radiomics-based classification is robust enough to translate beyond a single research environment.
The fully automatic pipeline, from segmentation through feature extraction to aggressiveness scoring, requires no radiologist input during operation. This makes it suitable for environments with limited specialist access and could support radiologists by highlighting lesions requiring closer attention in high-volume settings.
The system's output could be integrated with the existing PI-RADS scoring system used clinically for mpMRI interpretation, providing a quantitative aggressiveness score alongside the radiologist's qualitative assessment and potentially improving the predictive value of the combined evaluation.
Limitations include the relatively small patient cohort and the fact that training and testing were both conducted at Italian centers with similar patient populations. Future validation in more diverse populations and additional scanner types will be necessary before widespread clinical adoption.
The broader significance is demonstrating that a radiomics-based system trained at one institution can generalize to another with acceptable accuracy, which is a prerequisite for any AI tool deployed in real clinical workflows where patient data always comes from a variety of sources.