Novel Machine Learning Methods for Prostate Image Analysis

Abstract

Accurate registration and fusion of prostate magnetic resonance (MR) and transrectal ultrasound (TRUS) images are vital for improving the effectiveness of prostate cancer diagnosis. However, the inherent differences in dimensionality and modality between 3D MR and 2D ultrasound images pose significant challenges for direct alignment and fusion. This thesis proposes a noval machine learning–based workflow that systematically addresses these challenges through three key stages: 3D TRUS reconstruction, cross-modality registration, and fusion.

First, we develop a fully sensorless method for TRUS 3D reconstruction by leveraging the capability of modern ultrasound systems for dual-plane scanning. Specifically, we use modern image stitching and matching techniques to infer the relative positions of images in the sagittal direction. This allows us to determine the spatial relative positions of transverse images and construct the 3D volume without any external probe tracking device.

Second, to address the remaining modality gap between MR and ultrasound (US), we propose a modality translation–based registration framework. Unlike existing methods that transform one modality into another, we innovatively propose creating an abstract intermediate modality
and transform both modalities toward it. This reduces the difficulty of modality translation and increases our control over the intermediate representation. Specifically, we introduce two methods. The first, Partial Modality Translation (PMT), controls the shallow features in the image latent space to reduce texture discrepancies between modalities, mitigating the modality gap. While effective, PMT still retains excessive details irrelevant to registration. To address this, we further propose Anatomical Coherence Modality Translation (ACMT), which customizes the intermediate modality by explicitly preserving only boundary information. This removes excessive details that do not contribute to registration. Additionally, we adjusted the registration network to be anatomy-aware, directing the alignment focus toward the unified prostate interior, as achieved by ACMT. This leads to enhanced cross-modality registration performance.

Finally, we propose a training data-free, unsupervised image fusion method by extending Deep Image Prior (DIP) to fusion tasks. By treating the problem as an inverse one, our method enables DIP to achieve high-quality fusion with just a single input image pair, overcoming the issue of limited training data in clinical applications.

Comprehensive evaluations on clinical datasets demonstrate that our framework significantly improves the accuracy and consistency of prostate MR-US image registration and fusion.

Date of Award	9 Dec 2025
Original language	English
Awarding Institution	University of Bristol
Supervisor	Alin Achim (Supervisor) & Pui Anantrasirichai (Supervisor)