Radiological semantics discriminate clinically significant grade prostate cancer

Area of Science:

• Radiology
• Oncology
• Medical Imaging

Background:

• Diagnosing clinically significant prostate cancer is challenging due to disease multifocality and variability in interpreting multi-parametric Magnetic Resonance Imaging (mpMRI).
• Prostate Imaging and Data Systems (PI-RADS) scores often lack consensus, impacting the reliability of mpMRI in prostate cancer detection.
• This study investigates the utility of specific radiological traits (semantics) from mpMRI to improve the discrimination of clinically significant prostate cancer.

Purpose of the Study:

• To assess the ability of radiological semantic traits observed on mpMRI to discriminate clinically significant prostate cancer.
• To compare the diagnostic performance of semantic trait-based models against PI-RADS scores.
• To evaluate the reproducibility of semantic trait interpretation among radiologists.

Main Methods:

• Retrospective analysis of mpMRI studies from 103 prostate cancer patients with 167 targeted biopsies.
• Two radiologists independently scored 16 semantic traits (size, shape, border, lymphadenopathy) on mpMRI images.
• A linear classifier model using semantic traits was developed and validated using cross-validation; performance was compared to PI-RADS predictors.

Main Results:

• Several individual semantic features (ADC-intensity, Homogeneity, early-enhancement, T2-intensity, extraprostatic extension) showed univariate discrimination of high-grade Gleason scores (AUROC 0.54-0.68).
• A multivariable model using three semantic features (ADC-intensity, T2-intensity, enhancement homogenicity) achieved an average AUROC of 0.7.
• The PI-RADS based predictor had an average AUROC of 0.6, indicating lower discriminatory ability compared to the semantic model.

Conclusions:

• Semantic traits derived from mpMRI are associated with pathological findings and demonstrate higher inter-radiologist reproducibility.
• Multivariable models based on these semantic traits exhibit superior discriminatory ability for clinically significant prostate cancer compared to PI-RADS scores.
• These findings suggest that semantic analysis of mpMRI can enhance the accuracy and consistency of prostate cancer diagnosis.