Weakly supervised deep learning for multimodal MRI-TRUS registration: Toward assisting prostate biopsy guidance

Area of Science:

• Medical Imaging
• Machine Learning
• Prostate Cancer Diagnostics

Background:

• Accurate registration of magnetic resonance imaging (MRI) and transrectal ultrasound (TRUS) is crucial for effective prostate cancer detection and biopsy guidance.
• Current methods often require extensive labeled data, limiting their widespread clinical adoption.

Purpose of the Study:

• To develop and validate a weakly supervised learning framework for robust MRI-TRUS registration.
• To enhance prostate cancer detection and biopsy accuracy while minimizing the need for manual annotations.

Main Methods:

• A two-stage weakly supervised framework was implemented, incorporating an attention-enhanced U-Net for prostate segmentation and a residual-enhanced registration network (RERN) for image alignment.
• Models were trained and evaluated on public and clinical MRI-TRUS datasets, utilizing metrics such as Dice Similarity Coefficient (DSC) and Hausdorff distance (HD95).

Main Results:

• The segmentation model achieved high accuracy (DSC: MRI 0.9154, TRUS 0.9384) and generalizability to clinical data.
• The registration model demonstrated robust performance with HD95 of 10.18 mm (public) and 11.18 mm (clinical), and Target Registration Error (TRE) below 8.64 mm.
• Clinical validation showed preserved diagnostic integrity and enhanced diagnostic confidence among senior radiologists.

Conclusions:

• The proposed weakly supervised framework enables high-precision MRI-TRUS registration with minimal annotation requirements.
• The method ensures strong generalizability and clinical applicability, offering a valuable tool for prostate cancer management.