Machine learning-based analysis of a semi-automated PI-RADS v2.1 scoring for prostate cancer

Front Oncol 2022 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
The Subjectivity Problem in Prostate MRI Reporting

Prostate cancer is one of the most common cancers in men and the fifth leading cause of cancer-related death globally. Multiparametric MRI (mpMRI) has emerged as a powerful tool for detecting clinically significant prostate cancer before biopsy, improving cancer detection rates while reducing unnecessary procedures.

The standard framework for interpreting prostate MRI is PI-RADS version 2.1 (PI-RADS v2.1), a scoring system that guides radiologists in grading suspicious lesions from score 2 (unlikely to be clinically significant cancer) to score 5 (highly likely to be clinically significant cancer). PI-RADS was designed to standardize reporting across different radiologists and institutions.

Despite its purpose of standardization, PI-RADS interpretation remains significantly subjective. Studies have consistently shown poor to moderate inter-reader variability among radiologists -- the same MRI can receive different scores from different readers, leading to inconsistent clinical decisions. This variability is greatest for intermediate-risk lesions (scores 3 and 4), where the clinical stakes are highest.

An automated or semi-automated PI-RADS scoring system could address this variability by making lesion measurement and score assignment more objective and reproducible. Prior deep learning approaches had limited performance (AUC of 0.70-0.74) and were restricted to two-class classification only, leaving the challenge of four-class scoring unsolved.

TL;DR: PI-RADS scoring suffers from significant radiologist-to-radiologist variability, motivating the development of an automated scoring system that can objectively classify all four PI-RADS levels.
Pages 2-3
Patient Data and MRI Acquisition

The study included 59 men (mean age 65 plus or minus 8.5 years) with clinically suspected prostate cancer, retrospectively evaluated with institutional ethics board approval. The cohort was distributed across all four PI-RADS scores: 16 patients with score 2, 10 with score 3, 18 with score 4, and 15 with score 5.

All MRI was performed on a 1.5 Tesla scanner using standard prostate sequences. T2-weighted imaging (T2WI) provided detailed anatomical structure, while diffusion-weighted imaging (DWI) at five b-values (0, 500, 1000, 1500, and 2000 s/mm2) enabled calculation of apparent diffusion coefficient (ADC) maps that reflect tissue cellularity and cancer grade.

The study focused on the peripheral zone (PZ) of the prostate -- the region where 70-75% of prostate cancers originate. DWI at the highest b-value (b = 2000) was used as the primary sequence for lesion visualization, consistent with PI-RADS v2.1 guidelines, which specify DWI as the dominant sequence for peripheral zone scoring.

Lesion regions of interest were manually marked under supervision of an expert radiologist with more than 20 years of prostate MRI experience, following PI-RADS guidelines. These expert-verified annotations served as both the reference standard for comparison and the input for automated lesion measurement algorithms.

TL;DR: 59 prostate MRI cases across all four PI-RADS score levels were analyzed using T2WI, DWI, and ADC maps, with expert radiologist annotations as the reference standard.
Pages 3, 7, 8
The Semi-Automated Pipeline: Five Processing Steps

The proposed model processes each MRI case through a structured five-step pipeline. First, the prostate gland is segmented from DWI images using the Chan-Vese active contour model, which minimizes an energy function that distinguishes tissue inside and outside the prostate boundary. Morphological opening operations clean up any stray pixels.

Second, 3D image registration aligns the T2WI images to the DWI coordinate space using affine transformation with 12 degrees of freedom, allowing for scaling, shearing, translation, and rotation. This alignment is essential so that the same lesion region of interest can be applied consistently to all three image types (T2WI, DWI, and ADC).

Third, the prostate is divided into its two anatomical zones using a probabilistic atlas with partial volume correction: the peripheral zone and the transition zone. The atlas is built from training images and assigns each voxel a probability of belonging to each zone, with a partial volume algorithm resolving ambiguous boundary pixels.

Fourth, after expert lesion ROI marking, an automatic ellipse-fitting algorithm measures lesion maximum diameter and volume using a least-squares approach. These two measurements directly feed into PI-RADS v2.1 scoring criteria: PI-RADS uses a lesion size cutoff of 1.5 cm to distinguish score 4 from score 5, and requires lesion volume greater than 0.5 cc for clinically significant cancer.

TL;DR: The five-step pipeline performs prostate segmentation, image registration, zone segmentation, ROI marking, and automatic lesion measurement using ellipse-fitting to derive PI-RADS scores.
Pages 4-5
Lesion Measurement and Model Performance

The automated ellipse-fitting measurements closely matched expert manual measurements across all score levels. Mean lesion maximum diameters were 0.47 cm for score 2, 0.67 cm for score 3, 0.96 cm for score 4, and 1.45 cm for score 5 -- showing a clear, progressive increase in lesion size with increasing PI-RADS score that aligns with expected biology.

The semi-automated model correctly matched the radiologist's PI-RADS assessment in 50 out of 59 cases (approximately 85%), with a strong positive Pearson correlation of r = 0.94 (p less than 0.05). Score-specific correct classification rates were 93.75% for score 2, 90% for score 3, 83.34% for score 4, and 73.35% for score 5.

For four-class classification (score 2 vs. 3 vs. 4 vs. 5), the best-performing classifier was LDA, achieving 88.00% accuracy and AUC of 0.94. For two-class classification (low score 2-3 vs. high score 4-5), the linear SVM performed best with 93.20% accuracy and AUC of 0.99 -- indicating near-perfect discrimination between clinically significant and non-significant cancer categories.

These results surpassed prior automated approaches: a VGG-16 deep learning method achieved AUC of 0.74, and a CNN-based semi-automated approach reached 60% accuracy for two-class classification. The current model achieved these gains with only 59 patients by leveraging quantitative lesion measurements aligned with PI-RADS criteria rather than relying purely on pattern-learned features.

TL;DR: The semi-automated model matched radiologist PI-RADS scoring in 85% of cases with AUC of 0.94 for four-class scoring and 0.99 for the clinically critical low vs. high score classification.
Pages 5-6
Clinical Implications and Workflow Integration

The primary clinical value of the proposed model is in reducing inter-reader variability. By providing an objective, quantitative score derived from automated lesion measurements, the system could serve as a consistent baseline that radiologists -- especially less experienced ones -- can compare to their own assessment before finalizing a report.

The two-class (low vs. high score) classification is most directly clinically actionable. Distinguishing PI-RADS 2-3 (low suspicion, often no biopsy needed) from PI-RADS 4-5 (high suspicion, biopsy recommended) is the primary decision point in prostate cancer screening workflows. The model's AUC of 0.99 for this binary decision is strong evidence for clinical utility.

The model is particularly well-suited for non-expert radiologists who may not regularly evaluate prostate MRI. By automating lesion measurement and score calculation based on PI-RADS v2.1 criteria, it reduces the cognitive load of prostate MRI interpretation and the potential for scoring errors due to visual estimation of lesion size.

One limitation acknowledged by the authors is that the current model was validated only on peripheral zone lesions. Transition zone scoring in PI-RADS relies primarily on T2WI rather than DWI, which requires a different scoring approach. Future work should extend the model to cover the full prostate, and validate in larger multicenter cohorts before clinical deployment.

TL;DR: The model's 0.99 AUC for the low vs. high PI-RADS classification directly supports the primary clinical decision of whether to proceed to biopsy, particularly benefiting less experienced radiologists.
Page 7
Key Takeaways

This study demonstrates that a semi-automated PI-RADS v2.1 scoring model combining image processing with supervised machine learning can achieve high classification accuracy across all four PI-RADS score levels, outperforming previous deep learning approaches despite using a much smaller dataset.

The key innovation is grounding the model in PI-RADS v2.1's own scoring criteria -- specifically lesion maximum diameter and volume -- rather than using opaque end-to-end deep learning. This makes the model's decisions interpretable and aligned with clinical guidelines, an important advantage for radiologist acceptance and regulatory approval.

Classification accuracy of 88% for four-class scoring and 93% for two-class scoring, achieved with 59 patients, establishes strong preliminary evidence for this approach. The strong correlation (r = 0.94) between model scores and expert radiologist assessment suggests the model captures the same clinically meaningful features that radiologists evaluate.

Scaling to larger multicenter datasets, extending coverage to the transition zone, and automating lesion ROI marking are the natural next steps toward a fully automated, clinically deployable prostate cancer MRI reporting system.

TL;DR: A semi-automated PI-RADS v2.1 model using quantitative lesion measurements achieves 88% four-class and 93% two-class accuracy, outperforming prior AI methods while remaining interpretable and guideline-aligned.
Citation: Open Access, . Available at: PMC9730331.