Promise and challenges of heart chamber segmentation from non-contrast CT scans using contrastive unpaired image translation: a feasibility study

2026-06-22 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics, Image and Video Processing · Depth: Expert, quick

Summary

ChameleonNet, a novel framework, demonstrates the feasibility of segmenting heart chambers from non-contrast CT scans without requiring manual non-contrast annotations. This approach utilizes a Contrastive Unpaired Translation (CUT) network with decoupled contrastive learning (DCL) loss to synthesize non-contrast CT images from contrast CT scans. A Hausdorff distance loss-enhanced nnU-Net was then trained on 292 synthesized non-contrast scans, following the translation model's training on 35,538 contrast-enhanced and 37,197 non-contrast CT slices. Evaluation on 36 synthesized non-contrast images showed Dice similarity coefficients (DSC) ranging from 0.91 to 0.94 and 95th Hausdorff distances (HD95) from 3.63 mm to 5.74 mm across the four heart chambers. On 36 real non-contrast CT scans, Pearson correlations were between 0.82 and 0.93, with mean absolute percentage errors (MAPE) from 9.22% to 20.79%. While promising, volume errors, particularly for the left and right ventricles, indicate the need for further refinement before clinical deployment.

Key takeaway

For AI Scientists developing medical image analysis tools, this study suggests a viable path to overcome annotation scarcity for non-contrast CT. You can leverage unpaired image translation to synthesize necessary training data, reducing manual labeling efforts. However, carefully validate your models on real-world data, as observed volume errors, especially for ventricles, highlight the need for robust accuracy before clinical integration. Prioritize refinement to ensure reliability.

Key insights

Unpaired image translation can enable deep learning segmentation of heart chambers from non-contrast CT without direct non-contrast annotations.

Principles

• Synthesize target images from source images.
• Use decoupled contrastive learning for translation.
• Validate models on both synthetic and real data.

Method

ChameleonNet uses a CUT network with DCL loss to synthesize non-contrast CT from contrast CT. An nnU-Net, enhanced with Hausdorff distance loss, then segments four heart chambers from these synthesized images.

In practice

• Apply unpaired translation to other medical imaging tasks.
• Reduce reliance on scarce expert annotations.
• Evaluate segmentation models with volume agreement metrics.

Topics

• Heart Chamber Segmentation
• Non-contrast CT
• Contrastive Unpaired Translation
• Deep Learning
• Medical Imaging
• ChameleonNet

Best for: Computer Vision Engineer, AI Scientist, Research Scientist, Machine Learning Engineer

Related on AIssential

• Synthesize Realistic 3D Medical Images at Scale to Ship Pre‑Trained Models — NVIDIA's MAISI and NV-Generate frameworks provide scalable, open-source 3D medical image synthesis to overcome data bottlenecks.
• Evidential learning driven Breast Tumor Segmentation with Stage-divided Vision-Language Interaction — TextBCS enhances breast tumor segmentation in MRI by combining text prompts with evidential learning to address low contrast and blurred boundaries.
• When is 3D Worth It? A Resource-Performance Frontier for CNNs and Transformers in Lung CT — Input dimensionality significantly impacts model performance and stability in volumetric medical imaging.
• CA-GCL: Cross-Anatomy Global-Local Contrastive Learning for Robust 3D Medical Image Understanding — CA-GCL uses global-local contrastive learning and clinical-aware text augmentation to prevent representation collapse in 3D medical FVLP.
• Cross-Source Supervision for Bone Infection Segmentation in Dual-Modality PET-CT — Decoupled dual-source learning with PET-CT fusion improves bone infection segmentation by accommodating diverse clinical annotation philosophies.
• Unified MRI Brain Image Translation via Hierarchical Tumor Structure Comparison — HTSCGAN improves multi-modal MRI brain image translation by integrating hierarchical tumor structural information using a GAN with specific modules and loss functions.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.