TL;DR
This paper introduces a novel radiomics model, XGB-AS, that uses magnetic resonance imaging (MRI) to non-invasively identify prostate cancer patients suitable for active surveillance (AS). This approach aims to reduce the need for frequent, invasive biopsies, thereby minimizing patient discomfort and associated risks.
What Problem Does This Paper Address?
Prostate cancer (PCa) management often involves a delicate balance between aggressive treatment and active surveillance (AS). AS is a primary strategy for low- or favorable-intermediate-risk PCa, involving regular monitoring to detect progression. However, current AS protocols mandate annual prostate biopsies, which are invasive, painful, and carry risks of bleeding and infection. Furthermore, the visual assessment of MRI scans for AS candidacy, while improving, is subjective and prone to inter-radiologist variability. This study addresses the need for a non-invasive, objective method to accurately identify PCa patients who would benefit from AS, thereby avoiding unnecessary biopsies and their associated complications.
Methods
The researchers developed and validated five radiomics models: eXtreme Gradient Boosting (XGBoost), logistic regression (LR), random forest (RF), adaptive boosting (AdaBoost), and decision tree (DT). These models were trained using radiomic features extracted from T2-weighted MRI images. The process involved:
- Image Preprocessing: Standardized preparation of MRI scans.
- Feature Extraction: Extraction of 1,595 radiomic features, including shape, texture, histogram, and filter-based features, from the region of interest (ROI) of prostate lesions.
- Feature Selection: A two-step process using a t-test to identify highly differentiated features between AS and non-AS candidates, followed by LASSO regression with five-fold cross-validation for further selection.
- Model Development: Construction of classification models using the selected features.
- Validation: A rigorous three-fold cross-center validation was employed, where different combinations of hospitals served as training and external validation cohorts to ensure generalizability. Performance was evaluated using AUC, accuracy (ACC), sensitivity (SEN), and specificity (SPE).
Datasets
The study utilized a multicenter dataset comprising 956 PCa patients from six hospitals. These patients underwent 3.0-T MRI scans prior to prostate biopsy. The data included complete biopsy reports and clinical diagnostic records necessary for determining AS candidacy based on the National Comprehensive Cancer Network (NCCN) guidelines. No new benchmark datasets were introduced; rather, existing patient data from multiple institutions were aggregated and analyzed.
Key Insights and Ideas
- Radiomics as a Non-Invasive Tool: The study demonstrates the significant potential of radiomics, derived from standard T2W MRI sequences, to act as a non-invasive tool for AS candidate identification.
- XGBoost's Superiority: Among the tested models, the XGBoost-based classifier (XGB-AS) consistently outperformed others in terms of overall performance metrics.
- Addressing Ambiguous Cases: The XGB-AS model showed promise in identifying AS candidates among patients with discordant assessments between PI-RADS and NCCN guidelines, a common challenge in clinical practice.
- Feature Significance: Specific radiomic features, particularly original shape features like sphericity and various wavelet and LBP-3D features, were found to be crucial in distinguishing AS candidates from non-AS candidates. The finding that AS candidates' lesions exhibit more regular shapes is particularly insightful.
Key Results
The XGB-AS model achieved a mean AUC of 0.803, ACC of 0.693, SEN of 0.668, and SPE of 0.841 across the cross-validation. In external validation, XGB-AS-1 demonstrated a significantly higher AUC (0.851) compared to the PI-RADS assessment (0.697) for identifying AS candidates. The model also showed promising performance in subgroup analyses:
- Ambiguous Cases: XGB-AS-1 achieved an ACC of 67.9% in distinguishing AS candidates from ambiguous cases (discordance between PI-RADS and NCCN guidelines).
- Intermediate-Risk Group: The model achieved an ACC of 67.1% in identifying AS candidates within the intermediate-risk group (FIR and UFIR patients).
Limitations and Future Work
The authors acknowledge several limitations:
- Retrospective Design: The study was retrospective, and future work should involve prospective validation with larger datasets.
- Manual Segmentation: The manual segmentation of ROIs was time-consuming; an automated segmentation algorithm would improve efficiency.
- Ambiguous/Intermediate-Risk Case Numbers: The relatively small number of ambiguous and intermediate-risk cases limited the ability to train and validate a deep-learning model for these specific subgroups.
Future research should focus on developing more advanced models, potentially deep learning-based, with larger datasets to further improve the accuracy in stratifying ambiguous and intermediate-risk patients.
Summary updated