Prostate MRI/TRUS Image Registration Algorithm Based on Weak Supervision

Prostate cancer is an important global public health issue, being the second most prevalent male cancer worldwide. With the popularization of medical imaging equipment and the improvement of imaging level, prostate histopathological image analysis has become an important technical means of cancer diagnosis and treatment. In this paper, a new prostate 3D MRI(Magnetic Resonance Imaging) and TRUS(Transrectal Ultrasonography) registration cascade network is proposed. The network consists of three sub-networks, including affine registration sub-network (A-Net), deformable sub-network (D-Net and SD-Net). All subnetworks are trained with weak supervision to avoid problems such as difficulty in obtaining Ground Truth and determining similarity measure. The network uses a convolutional module with fusion residual connection, dense connection and depth separability, which is conducive to multi-scale feature extraction and fusion. Various attention mechanisms are integrated into the network to enhance training efficiency and ensure accurate registration. Experimental results on The Cancer Imaging Archive (TCIA) dataset demonstrate that the proposed method outperforms existing techniques, effectively fusing MRI and TRUS images for guiding prostate puncture surgery. TRE is 1.61±0.862mm, DSC is 0.912±0.037, and the average error of the final image fusion guided prostate puncture is 2.14±0.27mm. These results affirm the clinical significance and practical value of the designed medical image registration network for image-guided biopsies.