Spectral augmentation for heart chambers segmentation on conventional contrasted and unenhanced CT scans: an in-depth study

Abstract

Recently, machine learning has outperformed established tools for automated segmentation in medical imaging. However, segmentation of cardiac chambers still proves challenging due to the variety of contrast agent injection protocols used in clinical practice, inducing disparities of contrast between cavities. Hence, training a generalist network requires large training datasets representative of these protocols. Furthermore, segmentation on unenhanced CT scans is further hindered by the challenge of obtaining ground truths from these images. Newly available spectral CT scanners allow innovative image reconstructions such as virtual non-contrast (VNC) imaging, mimicking non-contrasted conventional CT studies from a contrasted scan. Recent publications have demonstrated that networks can be trained using VNC to segment contrasted and unenhanced conventional CT scans to reduce annotated data requirements and the need for annotations on unenhanced scans. We propose an extensive evaluation of this statement. We undertake multiple trainings of a 3D multi-label heart segmentation network with (HU-VNC) and without (HUonly) VNC as augmentation, using decreasing training dataset sizes (114, 76, 57, 38, 29, 19 patients). At each step, both networks are tested on a multi-vendor, multi-centric dataset of 122 patients, including different protocols: pulmonary embolism (PE), chest-abdomen-pelvis (CAP), heart CT angiography (CTA) and true non-contrast scans (TNC). An in-depth comparison of resulting Dice coefficients and distance metrics is performed for the networks trained on the largest dataset. HU-VNC-trained on 57 patients significantly outperforms HUonly trained on 114 regarding CAP and TNC scans (mean Dice coefficients of 0.881/0.835 and 0.882/0.416, respectively). When trained on the largest dataset, significant improvements in all labels are noted for TNC and CAP scans (mean Dice coefficient of 0.882/0.416 and 0.891/0.835, respectively). Adding VNC images as training augmentation allows the network to perform on unenhanced scans and improves segmentations on other imaging protocols, while using a reduced training dataset.