MRI-based prostate cancer classification using 3D efficient capsule network.

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Oncology

Background:

• Prostate cancer (PCa) risk stratification relies on biopsy, an invasive procedure.
• Magnetic resonance imaging (MRI) offers non-invasive characterization but faces interpretation variability.
• Gleason score (GS) is crucial for PCa risk but currently requires invasive assessment.

Purpose of the Study:

• To develop a 3D Efficient Capsule Network (CapsNet) for non-invasive prediction of PCa risk using T2-weighted (T2W) MRI.
• To overcome limitations of Convolutional Neural Networks (CNNs) in encoding spatial information for improved robustness.
• To stratify PCa risk based on quantitative MRI analysis.

Main Methods:

• Utilized 3D CNN modules for spatial feature extraction and primary capsule layers for vector encoding.
• Integrated fully-connected capsule layers (FC Caps) to create a deeper hierarchy for PCa grading.
• Employed a novel dynamic weighted margin loss function to address data imbalance.
• Evaluated the method on a public dataset of 976 PCa T2W MRI scans.

Main Results:

• The 3D Efficient CapsNet achieved high performance in classifying PCa risk, with AUCs up to 0.83 for low vs. high grade.
• The model outperformed state-of-the-art radiomics and deep learning methods in PCa risk stratification.
• A weighted Cohen's Kappa score of 0.41 indicated moderate agreement with ground truth PCa risks.

Conclusions:

• A novel 3D Efficient CapsNet demonstrates feasibility for non-invasive PCa risk stratification using T2W MRI.
• This tool has the potential to personalize PCa treatment and decrease the need for invasive biopsies.
• The developed method offers a promising non-invasive approach for assessing PCa risk from MRI data.