DEFORMABLE MRI TO TRANSRECTAL ULTRASOUND REGISTRATION FOR PROSTATE INTERVENTIONS WITH SHAPE-BASED DEEP VARIATIONAL AUTO-ENCODERS

Prostate cancer is one of the most prevalent cancers in men, where diagnosis is confirmed through biopsies analyzed with histopathology. A diagnostic T2-w MRI is often registered to intra-operative transrectal ultrasound (TRUS) for effective targeting of suspicious lesions during image-guided biopsy procedures or needle-based therapeutic interventions such as brachytherapy. However, this process remains challenging and time-consuming in an interventional environment. The present work proposes an automated 3D deformable MRI to TRUS registration pipeline that leverages both deep variational auto-encoders with a non-rigid iterative closest point registration approach. A convolutional FC-ResNet segmentation model is first trained from 3D TRUS images to extract prostate boundaries during the procedure. Matched MRI-TRUS 3D segmentations are then used to generate a vector representation of the gland's surface mesh between modalities, used as input to a 10-layer dense variational autoencoder model to constrain the predicted deformations based on a latent representation of the deformation modes. At each iteration of the registration process, the warped image is regularized using the autoencoder's reconstruction loss, ensuring plausible anatomical deformations. Based on a 5-fold cross-validation strategy with 45 patients undergoing HDR brachytherapy, the method yields a Dice score of 85.0 +/- 2.6 with a target registration error of 3.9 +/- 1.4mm, with the proposed method yielding results outperforming the state-of-the-art, with minimal intra-procedural disruptions.