Automating prostate biopsy guidance: A robust CNN approach for non-rigid 3D/3D MR-TRUS image registration

Fusing transrectal ultrasound (TRUS) and magnetic resonance (MR) images has significantly improved the accuracy of prostate cancer detection during targeted biopsies. However, automatic MR-TRUS registration remains challenging due to the large anatomical and appearance differences between the two imaging modalities. This study introduces a fully automatic, weakly supervised deep learning (DL) approach for predicting dense displacement fields (DDFs) between 3D MR and 3D TRUS images without requiring segmentation images. The proposed method consists of two main stages: (1) a dl-based preprocessing step that aligns the prostate centroids of the MR and TRUS images to improve initialization, and (2) a UNet-inspired registration network (RegResUNet) combined with a spatial transformer layer (STL) that directly predicts voxel-level DDFs from the aligned 3D input images. The network is trained using weak supervision based on anatomical masks, enabling accurate registration without ground truth DDFs. Experimental results demonstrate that the proposed method achieves substantial improvements over rigid baseline registration, with an average surface registration error (SRE) of 0.97 ± 0.85 mm and an average Dice similarity coefficient (DSC) of 0.93 ± 0.02. Notably, the proposed network outperforms several state-of-the-art non-rigid registration models while maintaining computational efficiency. The whole pipeline eliminates the need for intermediate manual and time-consuming segmentation steps. The automatic and robust capabilities of the proposed approach, combined with its short inference time, highlight its potential for clinical use in prostate cancer interventions, reducing human factors, ensuring consistent results, and contributing to improved patient comfort by minimizing procedure time.