Abstract: A combination of texture features and machine learning-based analysis of ADC maps could predict PCa GG upgrading from biopsy to radical prostatectomy non-invasively with satisfactory predictive efficacy.
Introduction
In the management of low-risk prostate cancer (PCa), Active Surveillance (AS) is a preferred strategy to avoid the morbidity associated with radical prostatectomy or radiation. However, the success of AS depends on the accuracy of the initial needle biopsy. A significant clinical challenge is pathological upgrading, where the final pathology after surgery reveals a higher Grade Group (GG) than initially detected by biopsy. This discrepancy occurs because biopsies sample only a fraction of the gland, potentially missing more aggressive tissue. This study explores a non-invasive machine learning framework using MRI "texture features" to predict this upgrading, identifying patients who may require immediate intervention rather than surveillance.
Methods
The study utilized a radiomics-based approach focusing on Apparent Diffusion Coefficient (ADC) maps, which quantify water diffusion—a proxy for tissue density and cellularity.
- Feature Extraction: Researchers used MaZda software to extract 94 texture features from 3D tumor volumes, including histograms, absolute gradients, and Gray-Level Co-occurrence Matrix (GLCM) features.
- Feature Selection: The LASSO (Least Absolute Shrinkage and Selection Operator) algorithm was applied to reduce dimensionality, selecting the 6 most predictive features to prevent overfitting.
- Machine Learning Classifiers: Four models were evaluated: Random Forests (RF), Decision Trees (DT), Support Vector Machines (SVM), and the Nearest Neighbor Algorithm (NNA).
Datasets
The retrospective study included 59 patients from Tongji Hospital (2016–2020) who had biopsy-proven GG 1 or 2 cancer and underwent subsequent radical prostatectomy. The cohort was split into a training set (n=48) and a test set (n=11). All imaging was performed on a consistent 3.0 T MRI scanner.
Results and Key Insights
The analysis revealed that the most predictive features (variance, skewness, kurtosis, etc.) were indicators of tumor heterogeneity. A "messier" texture on the ADC map strongly correlated with hidden, higher-grade cancerous cells.
- Model Performance: In the training cohort, the Nearest Neighbor Algorithm (NNA) achieved the highest Area Under the Curve (AUC) of 0.872 (sensitivity 96.7%), followed by the SVM with an AUC of 0.861.
- Clinical Utility: All machine learning models outperformed traditional clinical markers like PSA levels. While performance slightly decreased in the small test cohort (NNA AUC 0.717), the results remained statistically significant, validating the predictive power of texture analysis.
Discussion and Limitations
This study provides a preliminary proof-of-concept for using radiomics to refine patient selection for active surveillance. However, several limitations must be addressed:
- Sample Size: The small cohort (n=59) necessitates larger, multi-center validation.
- Selection Bias: The model currently only applies to patients with visible lesions on ADC maps, excluding "MRI-invisible" cancers.
- Anatomical Zonal Differences: Future work should distinguish between the peripheral and transition zones of the prostate, as cancer behavior varies by location.
Despite these hurdles, the integration of machine learning and quantitative MRI analysis represents a promising step toward more personalized and accurate prostate cancer management.
Summary updated