Work

Repeatability of prostate MRI-based radiomics features: The influence of lesion segmentation volume

Year: 2025

Type: article

Abstract: MRI radiomics is an effective tool for assessing prostate cancer. The volume of the region of interest (ROI) in prostate cancer lesions varies among patients, which may affect the intra-observer repeatability of radiomics features. The aim of this study was to evaluate the impact of varying prostate lesion ROI volumes on the intra-observer repeatability of radiomics features in prostate MRI. This study involved 337 men from the ProstateX dataset, all with a PI-RADS score of 3 or higher who underwent MRI-guided targeted biopsy. Imaging was performed using a 3 T MRI with T2-weighted imaging (T2WI), diffusion-weighted imaging (DWI), and ADC maps. Prostate lesions were manually segmented by three qualified observers, and the Sørensen-Dice coefficient assessed the similarity of the segmentations. Lesions were categorized into five groups based on average ROI volume: A (0–500 mm³), B (500–1000 mm³), C (1000–1500 mm³), D (1500–2000 mm³), and E (> 2000 mm³). Intra-class correlation coefficients (ICC) evaluated intra-observer repeatability, while Kendall correlation analysis explored the relationship between repeatability and ROI volume. After preprocessing, radiomics features were extracted from T2WI and ADC maps for further analysis, with P < 0.05 considered statistically significant. Group E had the highest repeatability of radiomics features on T2WI (83.1 % with ICC ≥ 0.75) and ADC map (81.3 % with ICC ≥ 0.75). In contrast, Group A had the lowest repeatability on T2WI (46.6 % with ICC ≥ 0.75) and ADC map (28.3 % with ICC ≥ 0.75). Furthermore, ROI volume showed a significant positive correlation with repeatability (p < 0.01). The intra-observer repeatability of radiomics features is significantly positively correlated with the volume of the prostate lesion ROI, highlighting the importance of considering ROI volume as a potential confounding factor in study design to enhance the reliability of radiomics analyses. • Intra-observer repeatability improves with larger lesion volumes. • ADC-derived features are more affected by lesion volume than T2-weighted imaging features. • First-order features show stable repeatability across different volumes, unlike shape and texture features. • Consider lesion volume impact in future radiomics research.

Source: European Journal of Radiology Artificial Intelligence

DOI: 10.1016/j.ejrai.2025.100008