Comparison of Prostate MRI Lesion Segmentation Agreement Between Multiple Radiologists and a Fully Automatic Deep...

Area of Science:

• Radiology
• Artificial Intelligence
• Medical Imaging Analysis

Background:

• Deep learning models, such as U-Net, are emerging tools for analyzing prostate MRI.
• These models aim to predict clinically significant prostate cancer (sPC) and segment lesions.

Purpose of the Study:

• To compare the segmentation agreement of U-Net with manual segmentations by multiple radiologists for prostate MRI lesions.
• To assess the performance of AI in lesion segmentation against human expert variability.

Main Methods:

• Retrospective analysis of 165 patients with suspected sPC undergoing MRI and biopsy.
• Generation of segmentations by three radiologists and a U-Net model.
• Calculation of Dice coefficients to quantify per-lesion agreement between segmentations.

Main Results:

• Manual segmentations showed moderate agreement (Dice coefficient 0.48-0.52), indicating difficulty in outlining lesions.
• U-Net segmentations had significantly lower agreement (Dice coefficient 0.22) compared to manual segmentations.
• Differences in agreement persisted after adjusting for lesion size and type.

Conclusions:

• Human inter-rater agreement for prostate MRI lesion segmentation is moderate, setting a benchmark for AI.
• U-Net's lower Dice coefficients suggest areas for improvement in AI segmentation, potentially by focusing on lesion cores.
• While AI's predictive performance is comparable, segmentation quality metrics like Dice coefficient are valuable secondary measures.