[Paper Review] Pseudo-Label Guided Multi-Contrast Generalization for Non-Contrast Organ-Aware Segmentation

Ho Hin Lee, Yucheng Tang+8 more|arXiv (Cornell University)|May 12, 2022

Advanced X-ray and CT Imaging4 citations

TL;DR

This paper proposes a pseudo-label guided unsupervised method for non-contrast CT (NCCT) organ segmentation by leveraging pairwise contrast-enhanced CT (CECT) context as teacher supervision, avoiding generative adversarial networks. It improves organ-aware boundary detection through intensity correlation modeling and achieves state-of-the-art performance, with +3.98% Dice on internal multi-organ and +8.00% on external aorta segmentation benchmarks.

ABSTRACT

Non-contrast computed tomography (NCCT) is commonly acquired for lung cancer screening, assessment of general abdominal pain or suspected renal stones, trauma evaluation, and many other indications. However, the absence of contrast limits distinguishing organ in-between boundaries. In this paper, we propose a novel unsupervised approach that leverages pairwise contrast-enhanced CT (CECT) context to compute non-contrast segmentation without ground-truth label. Unlike generative adversarial approaches, we compute the pairwise morphological context with CECT to provide teacher guidance instead of generating fake anatomical context. Additionally, we further augment the intensity correlations in 'organ-specific' settings and increase the sensitivity to organ-aware boundary. We validate our approach on multi-organ segmentation with paired non-contrast & contrast-enhanced CT scans using five-fold cross-validation. Full external validations are performed on an independent non-contrast cohort for aorta segmentation. Compared with current abdominal organs segmentation state-of-the-art in fully supervised setting, our proposed pipeline achieves a significantly higher Dice by 3.98% (internal multi-organ annotated), and 8.00% (external aorta annotated) for abdominal organs segmentation. The code and pretrained models are publicly available at https://github.com/MASILab/ContrastMix.

Motivation & Objective

• To address the challenge of organ boundary ambiguity in non-contrast CT (NCCT) scans due to lack of contrast enhancement.
• To enable accurate organ segmentation in NCCT without requiring expensive human-annotated ground-truth labels.
• To improve sensitivity to organ-specific anatomical boundaries through multi-contrast context modeling.
• To develop a generalizable, unsupervised framework that transfers knowledge from paired CECT scans to NCCT segmentation.
• To validate performance on both internal multi-organ and external independent aorta segmentation cohorts.

Proposed method

• Uses paired NCCT and CECT scans to compute morphological context as pseudo-label guidance, replacing generative adversarial training.
• Leverages pairwise CECT scans to extract structural context that guides NCCT segmentation in an unsupervised manner.
• Enhances intensity correlation learning within organ-specific regions to improve boundary sensitivity.
• Applies multi-contrast generalization by modeling spatial and intensity relationships across contrast types.
• Employs a self-supervised training pipeline where CECT-derived pseudo-labels act as teacher signals for NCCT student network.
• Utilizes five-fold cross-validation and full external validation on an independent NCCT cohort for robustness evaluation.

Experimental results

Research questions

• RQ1 Can multi-contrast CECT context be effectively used to guide unsupervised NCCT organ segmentation without ground-truth labels?
• RQ2 How does pseudo-label guidance from CECT pairs compare to generative adversarial approaches in preserving organ boundaries?
• RQ3 To what extent does modeling organ-specific intensity correlations improve segmentation accuracy on NCCT?
• RQ4 Does the proposed method generalize to independent external NCCT cohorts, particularly for challenging structures like the aorta?
• RQ5 Can the method achieve state-of-the-art performance in NCCT organ segmentation under fully unsupervised settings?

Key findings

• The method achieves a 3.98% higher Dice score on internal multi-organ segmentation compared to the current state-of-the-art in fully supervised settings.
• On an independent external aorta segmentation cohort, the method achieves an 8.00% higher Dice score than the current SOTA.
• The model demonstrates strong generalization to unseen NCCT scans, validated through full external testing.
• The use of CECT-derived morphological context as pseudo-labels outperforms generative adversarial training in boundary preservation.
• The approach effectively learns organ-aware representations by modeling intensity correlations in organ-specific regions.
• The code and pretrained models are publicly released, enabling reproducibility and further research.