Radiomics and machine learning of multisequence multiparametric prostate MRI: Towards improved non-invasive prostate cancer characterization

PLoS One 2019 Machine Learning 6 Explanations View Original
Original Paper (PDF)

Unable to display PDF. Download it here or view on PMC.

Plain-English Explanations
Pages 1-2
Non-Invasive Gleason Score Prediction

Gleason score is the most important marker for estimating prostate cancer aggressiveness, determined by examining glandular architecture patterns in biopsy or surgical specimens under a microscope. However, standard biopsy methods fail to accurately reflect the true Gleason score in 30% to 50% of patients, due to sampling error in a heterogeneous tumor.

Diffusion-weighted MRI (DWI) has shown significant potential for non-invasively characterizing prostate cancer. Most studies measure only simple average values (first-order statistics) from the commonly derived apparent diffusion coefficient (ADC) map, overlooking the rich spatial texture information encoded in how pixel intensities vary across the tumor region.

More advanced DWI signal modeling using the kurtosis model -- which accounts for non-Gaussian water diffusion behavior at high magnetic field gradients -- provides a more accurate description of tissue microstructure than the standard monoexponential model, but its texture features had not previously been evaluated for Gleason score prediction.

This study aimed to rigorously evaluate a large number of texture and statistical features from multiple MRI sequences, identify the best combinations for predicting Gleason score non-invasively, and provide the research community with fully open data and code to enable reproducibility.

TL;DR: This study systematically evaluated thousands of radiomic texture features from multiple MRI sequences to identify which combinations best predict prostate cancer Gleason score without biopsy.
Pages 3-5
High-Quality Multiparametric MRI Acquisition

62 patients with histologically confirmed prostate cancer scheduled for prostatectomy were scanned at 3 Tesla MRI using three complementary imaging sequences: T2-weighted anatomical imaging (T2w), diffusion-weighted imaging acquired at 12 different b-values ranging from 0 to 2000 s/mm2, and T2 relaxation mapping.

The DWI data was mathematically modeled using two approaches: the standard monoexponential model (producing ADCm maps) and the more sophisticated kurtosis model (producing ADCk and K parameter maps), which better captures non-Gaussian water diffusion behavior in complex, heterogeneous tissue.

Cancer regions were manually outlined on each MRI by an expert, guided by direct comparison with whole-mount prostatectomy sections -- the gold standard histopathology preparation where the entire removed prostate is sliced and examined. This process ensured highly accurate tumor delineation for each imaging modality.

T2-weighted images underwent intensity standardization to a common scale using histogram alignment, correcting for the inherent non-quantitative nature of MRI signal intensity that varies between patients and scanning sessions.

TL;DR: Carefully acquired high-quality MRI data with histopathology-guided tumor delineation ensured that the feature extraction and machine learning analysis were anchored to accurate ground truth.
Pages 6-9
7,105 Features from Eight Texture Methods

From the five types of parametric maps (T2w, ADCm, ADCk, K, T2), a total of 7,105 unique texture features were extracted per tumor using eight different texture analysis methods, each probing a different aspect of how pixel intensities are spatially organized within the tumor region.

The methods spanned multiple categories: the gray-level co-occurrence matrix (GLCM) captures statistical relationships between neighboring pixels; Gabor filters detect edges and spatial frequency patterns at different orientations; Local Binary Patterns (LBP) encode local intensity transformation patterns; and Zernike and Hu moments capture the overall spatial mass distribution of image intensity.

Each method was applied using multiple sliding window sizes -- from 3 to 35 pixels depending on image resolution -- to capture texture at different spatial scales simultaneously, since the optimal scale for distinguishing tissue types is not known a priori and varies by method and data.

Additionally, 18 first-order statistical features (mean, standard deviation, quartiles, skewness, kurtosis, etc.) were computed from each image type as a comparison baseline, representing the simpler analysis approach used in most prior studies.

TL;DR: An extensive feature extraction pipeline computed over 7,000 texture and statistical measures from five MRI parameter maps using eight different mathematical analysis methods.
Pages 9-10
Rigorous Cross-Validation to Prevent Overfitting

With 7,105 candidate features for only 100 lesions, preventing the model from simply memorizing the training data (overfitting) was the central methodological challenge. The solution was regularized logistic regression, which penalizes overly complex models, tested with both L1 regularization (which also performs feature selection) and L2 regularization (which weights all features).

Feature selection was performed by ranking features by their individual AUC (area under the ROC curve) and selecting only the top 1% -- approximately 12 to 16 features per image type -- ensuring only the most discriminative features entered the final model.

Model performance was estimated using leave-pair-out cross-validation (LPOCV), an exhaustive strategy where every possible pair of lesions is held out in turn as a test set while all remaining lesions train the model. This provides an unbiased estimate of true generalization performance and is particularly appropriate for small datasets.

Crucially, the feature selection and hyperparameter tuning steps were performed entirely within each LPOCV training fold, so the test data never influenced any modeling decision -- a strict protocol that prevents the optimistic bias common in studies that perform feature selection on the full dataset before cross-validation.

TL;DR: An exhaustive leave-pair-out cross-validation approach with within-fold feature selection ensured unbiased performance estimates despite the large number of candidate features relative to training samples.
Pages 11, 13, 14
Best Results Combining Diffusion and T2w Features

For individual imaging sequences, the best single-image performance came from T2-weighted imaging: a GLCM homogeneity feature achieved an AUC of 0.84 as a single feature, and the T2w top-1% feature model achieved AUC = 0.82, outperforming all DWI-derived models on their own.

Combining features across imaging types substantially improved performance. The best overall model used the top 1% features from ADCm, kurtosis parameter K, and T2w combined, achieving an AUC of 0.88 (95% CI: 0.82-0.95) for distinguishing low-risk (Gleason 3+3) from higher-risk (Gleason greater than 3+3) tumors.

The kurtosis-derived features from DWI (particularly the kurtosis parameter K) provided improvement over monoexponential ADC features alone, confirming that the more advanced diffusion model captures additional biologically relevant information about tumor microstructure -- an important finding since most clinical practice only uses simpler ADC.

In contrast, T2 mapping features contributed little to the best-performing models, suggesting that T2 relaxation time maps -- while providing some information individually -- are largely redundant with T2w and DWI features in this context.

TL;DR: Combining selected texture features from T2-weighted imaging and kurtosis DWI parameters achieved the best Gleason score classification with an AUC of 0.88, outperforming any single imaging sequence alone.
Pages 15-17
Which Texture Methods Work Best and Why

Gabor filter features ranked among the top 1% for all imaging types, suggesting that edge and frequency patterns detected at specific spatial scales are universally informative across DWI and T2w images, likely reflecting structural boundaries between tumor glands and stroma.

Zernike and Hu moments -- which describe how image intensity is distributed spatially, analogous to the center of mass and spread of a distribution -- performed particularly well for kurtosis K and T2 maps. These methods had not previously been applied to prostate cancer MRI texture analysis, and the authors argue they are effective because tumor Gleason grade correlates with tissue heterogeneity patterns that manifest as different intensity mass distributions.

A noteworthy finding was that texture features consistently outperformed simple statistical features for T2w images, while for DWI-derived ADC maps the simpler statistical features (such as minimum ADC value) performed comparably to texture features. This suggests that the most useful information in T2w is spatial, while ADC intensity levels alone carry much of the relevant signal.

Study limitations include the relatively small cohort of 62 patients, the bias toward high Gleason score cases (80% of lesions), and the single-center, single-scanner design. Larger multi-center validation studies are needed before this approach can be considered for clinical translation.

TL;DR: Gabor and moment-based texture features were the most informative for Gleason score prediction, with different optimal methods emerging for different MRI image types -- highlighting the need for sequence-specific feature optimization.
Citation: Open Access, . Available at: PMC6613688.