[Paper Review] Atlas Based Segmentations via Semi-Supervised Diffeomorphic Registrations

Charles Huang, Masoud Badiei+6 more|arXiv (Cornell University)|Nov 23, 2019

Medical Image Segmentation Techniques36 references4 citations

TL;DR

This paper proposes a semi-supervised diffeomorphic registration framework that leverages an atlas with ground truth contours for a subset of organs-at-risk (OARs) and rough segmentations for others to generate accurate, full OAR segmentations in head and neck radiotherapy patients. By cascading rigid and deformable deformation blocks with variational regularization and pseudo-labeling, the method achieves a mean Dice score of 0.789 on a test set of 20 patient scans, outperforming state-of-the-art learning-based registration and matching supervised segmentation performance while enabling automated segmentation of all OARs.

ABSTRACT

Purpose: Segmentation of organs-at-risk (OARs) is a bottleneck in current radiation oncology pipelines and is often time consuming and labor intensive. In this paper, we propose an atlas-based semi-supervised registration algorithm to generate accurate segmentations of OARs for which there are ground truth contours and rough segmentations of all other OARs in the atlas. To the best of our knowledge, this is the first study to use learning-based registration methods for the segmentation of head and neck patients and demonstrate its utility in clinical applications. Methods: Our algorithm cascades rigid and deformable deformation blocks, and takes on an atlas image (M), set of atlas-space segmentations (S_A), and a patient image (F) as inputs, while outputting patient-space segmentations of all OARs defined on the atlas. We train our model on 475 CT images taken from public archives and Stanford RadOnc Clinic (SROC), validate on 5 CT images from SROC, and test our model on 20 CT images from SROC. Results: Our method outperforms current state of the art learning-based registration algorithms and achieves an overall dice score of 0.789 on our test set. Moreover, our method yields a performance comparable to manual segmentation and supervised segmentation, while solving a much more complex registration problem. Whereas supervised segmentation methods only automate the segmentation process for a select few number of OARs, we demonstrate that our methods can achieve similar performance for OARs of interest, while also providing segmentations for every other OAR on the provided atlas. Conclusions: Our proposed algorithm has significant clinical applications and could help reduce the bottleneck for segmentation of head and neck OARs. Further, our results demonstrate that semi-supervised diffeomorphic registration can be accurately applied to both registration and segmentation problems.

Motivation & Objective

• Address the clinical bottleneck in radiation oncology caused by time-consuming, manual segmentation of organs-at-risk (OARs) in head and neck CT scans.
• Overcome limitations of supervised deep learning models that require complete, expert-annotated OAR datasets, which are often unavailable due to high manual effort.
• Develop a semi-supervised registration framework that uses a single atlas with partial ground truth contours and rough segmentations to generate full OAR segmentations for all structures.
• Enable robust, automated segmentation across all OARs—even those not explicitly annotated in training—by leveraging deformable image registration with diffeomorphic constraints.
• Demonstrate clinical utility by achieving performance comparable to manual and supervised segmentation while solving a more complex registration problem than prior learning-based methods.

Proposed method

• The method cascades rigid and deformable deformation blocks to progressively align a patient CT image to an atlas image, ensuring diffeomorphic (smooth, invertible) transformations.
• It uses a variational formulation with regularization terms and Gaussian smoothing of velocity fields to prevent overfitting and maintain anatomical plausibility.
• The network employs dense convolutional blocks in a localization network to improve spatial feature learning and deformation accuracy.
• Semi-supervised pseudo-labeling is applied during training: unlabeled OARs in the atlas are used to generate pseudo-segmentations, which are refined through iterative registration.
• The framework takes as input: a patient CT (F), an atlas CT (M), and atlas-space segmentations (S_A), and outputs patient-space segmentations for all OARs.
• Deformation fields are computed via integration of stationary velocity fields using a scaling and squaring method to preserve diffeomorphism.

Experimental results

Research questions

• RQ1 Can a semi-supervised learning-based registration framework achieve high-accuracy OAR segmentation in head and neck radiotherapy using only partial ground truth contours?
• RQ2 Does cascading multiple deformation blocks with variational regularization improve registration and segmentation performance compared to single-step or baseline learning-based methods?
• RQ3 How does the proposed method perform in the presence of image artifacts such as dental implants, which commonly degrade segmentation accuracy?
• RQ4 To what extent can a single atlas with limited ground truth contours and rough segmentations for other OARs produce full, accurate patient-level segmentations across all structures?
• RQ5 Can the method match or exceed the performance of supervised deep learning segmentation models while avoiding the need for complete, expert-annotated datasets?

Key findings

• The proposed method achieves a mean Dice score of 0.789 on a test set of 20 head and neck patient CT scans, significantly outperforming existing learning-based registration algorithms.
• The method demonstrates robustness to image artifacts such as dental implants, generating reasonable OAR contours even in regions obscured by metal artifacts.
• Performance is comparable to both manual and supervised deep learning segmentation methods, despite the model being trained on only partial ground truth data.
• The use of pseudo-labeling and variational regularization reduces overfitting, especially when using multiple deformation blocks in the cascade.
• The method outperforms traditional multi-atlas-based auto-segmentation (multi-ABAS) algorithms and matches state-of-the-art supervised segmentation performance.
• The framework enables full OAR segmentation across all atlas-defined structures, even those without explicit ground truth in training, by leveraging semi-supervised learning and diffeomorphic registration.