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

Accurate segmentation of heart chambers is vital for diagnosing cardiovascular diseases and planning radiation therapy. While contrast-enhanced CT scans provide clear anatomical detail, they are not always available or appropriate for every patient. Non-contrast CT scans are more accessible but lack the tissue differentiation required for precise manual segmentation. This study investigates whether deep learning can bridge this gap by synthesizing non-contrast CT images from contrast-enhanced scans to train a robust segmentation model.

The authors developed ChameleonNet, a two-stage framework. In the first stage, an image translation network—based on the Contrastive Unpaired Translation (CUT) architecture—converts contrast-enhanced CT scans into synthetic non-contrast images. To improve training efficiency, the authors introduced a decoupled contrastive learning (DCL) loss, which mitigates the negative-positive coupling effect common in standard contrastive learning. In the second stage, a modified nnU-Net segmentation model is trained on these synthetic images. This model incorporates a Hausdorff distance loss to better capture the spatial boundaries of the four heart chambers: the left atrium (LA), left ventricle (LV), right atrium (RA), and right ventricle (RV).

ChameleonNet successfully generated synthetic non-contrast images that maintained structural integrity while mimicking the intensity distribution of real non-contrast scans. The segmentation model achieved high Dice similarity coefficients (DSC) on synthesized images, ranging from 0.91 to 0.94 across the four chambers. When tested on real non-contrast CT scans, the model showed strong Pearson correlations with ground truth volumes (0.82 to 0.93). However, the mean absolute percentage error (MAPE) for volume estimation reached up to 20.79%, particularly for the ventricles, highlighting a performance gap between synthetic training data and real-world clinical scans.