A Multi-Task Deep Neural Network for Segmentation and Landmark Detection in Cardiac Computerized Tomography

Nicla Mandas1, Giulia Baldazzi2, Andrea Pitzus3, Giacomo Tarroni4, Danilo Pani3
1The Hadron Academy, IUSS, Pavia; DIEE, University of Cagliari, 2DIEE, University of Cagliari;, 3DIEE - University of Cagliari, 4City St George's, University of London


Abstract

Introduction: Multimodal bioimaging is increasingly recognized for its potential to integrate multiple types of information. This is particularly relevant in interventional cardiology, where structural imaging may be fused with electrophysiological data. Automating the preprocessing steps required for image alignment and registration is crucial to accelerate procedures in clinical settings. This study explores the feasibility of using a multi-task deep neural network for the automatic segmentation of the left ventricle (LV) from cardiac computerized tomography scans and the prediction of a landmark position required for image alignment. Methods: The proposed deep learning model is a multi-task network based on a 3D UNet, which simultaneously performs the segmentation of the LV and the localization of its apex. This model was trained and tested on the segmented images of the Multi-Modality Whole Heart Segmentation dataset, where the apex was manually annotated by an expert. Moreover, a 10-fold cross-validation was performed, with the dataset being partitioned into training, validation, and test sets following an 80/10/10 split. As performance parameters, the Dice score was adopted for the segmentation task, whereas the Euclidean distance for the landmark coordinates. Results: The model achieved consistent and robust performance in the segmentation task, while landmark prediction proved to be more challenging and exhibited greater variability. Across folds, the network achieved an average Dice score of 0.91 and an average Euclidean distance of 11.28mm for the segmentation and the landmark detection, respectively. Conclusions: This study proved the feasibility of using a multi-task network for the segmentation of the LV and the prediction of the apex coordinates. The proposed model achieved a high segmentation accuracy and a suitable landmark localization performance. These results suggest that, with some improvements, the proposed technique could be used as a preprocessing step when aligning the volumetric image of a cardiac chamber to another structure.