Deep adaptive registration of multi-modal prostate images

Area of Science:

• Medical Imaging
• Artificial Intelligence
• Deep Learning

Background:

• Deep learning significantly impacts cancer imaging, but multi-modal image registration for prostate cancer biopsy guidance faces challenges due to limited labeled data.
• Accurate registration of transrectal ultrasound (TRUS) and magnetic resonance (MR) images is crucial for improving prostate cancer diagnosis.

Purpose of the Study:

• To address the challenge of limited labeled data in multi-modal image registration for prostate cancer.
• To develop and evaluate a novel method for generating large amounts of transformations for network training.
• To implement a coarse-to-fine multi-stage registration technique for improved accuracy.

Main Methods:

• A new method for generating targeted transformations to augment training data for deep learning models.
• A multi-stage, coarse-to-fine registration approach to gradually align multi-modal prostate images (T2-weighted MR and 3D ultrasound).
• Evaluation of the proposed methods on a multi-modal prostate image registration task.

Main Results:

• Data generation significantly reduced registration error by up to 62%.
• The multi-stage coarse-to-fine registration achieved a mean surface registration error (SRE) of 3.66 mm, outperforming one-step registration (4.08 mm SRE).
• The initial mean SRE was 9.42 mm, demonstrating substantial improvement.

Conclusions:

• The developed data generation technique effectively enhances deep learning network training for image registration.
• The coarse-to-fine multi-stage registration method significantly improves accuracy in aligning multi-modal prostate images.
• These advancements hold promise for more precise image-guided prostate cancer biopsies.