[Paper Review] Automatic Segmentation of the Left Ventricle in Cardiac CT Angiography Using Convolutional Neural Network

Majd Zreik, Tim Leiner+4 more|arXiv (Cornell University)|Apr 19, 2017

Medical Imaging Techniques and Applications4 citations

TL;DR

This paper presents a two-stage deep learning method for automatic left ventricle (LV) segmentation in cardiac CT angiography (CCTA) using convolutional neural networks (CNNs). First, three CNNs detect a bounding box around the LV across axial, sagittal, and coronal planes; then, a voxel-wise CNN classifies structures within the box. The method achieves a mean Dice coefficient of 0.85 and a mean absolute surface distance of 1.1 mm, demonstrating high accuracy and feasibility for clinical use.

ABSTRACT

Accurate delineation of the left ventricle (LV) is an important step in evaluation of cardiac function. In this paper, we present an automatic method for segmentation of the LV in cardiac CT angiography (CCTA) scans. Segmentation is performed in two stages. First, a bounding box around the LV is detected using a combination of three convolutional neural networks (CNNs). Subsequently, to obtain the segmentation of the LV, voxel classification is performed within the defined bounding box using a CNN. The study included CCTA scans of sixty patients, fifty scans were used to train the CNNs for the LV localization, five scans were used to train LV segmentation and the remaining five scans were used for testing the method. Automatic segmentation resulted in the average Dice coefficient of 0.85 and mean absolute surface distance of 1.1 mm. The results demonstrate that automatic segmentation of the LV in CCTA scans using voxel classification with convolutional neural networks is feasible.

Motivation & Objective

• To develop an automatic, deep learning-based method for left ventricle segmentation in cardiac CT angiography (CCTA) scans.
• To overcome the limitations of manual segmentation, including time consumption and inter-observer variability.
• To leverage end-to-end learning with CNNs to eliminate manual feature engineering and improve segmentation accuracy.
• To evaluate performance on a diverse set of CCTA scans with clinical relevance, including apical regions and complex structures like papillary muscles.

Proposed method

• A two-stage approach is used: first, three separate CNNs detect the presence of the left ventricle in axial, sagittal, and coronal image planes to define a bounding box.
• Each CNN processes all slices in its respective plane to localize the LV, generating a 3D bounding box around the structure.
• Within the defined bounding box, a dedicated voxel classification CNN analyzes three orthogonal 48×48 image patches (axial, sagittal, coronal) centered on each voxel.
• The input to the voxel classifier consists of three image patches from orthogonal planes, enabling multi-view context for accurate tissue classification.
• The final segmentation is produced by the CNN’s softmax output layer, classifying each voxel as LV or non-LV based on learned features.
• The method is trained end-to-end using 50 scans for localization, 5 for segmentation, and 5 for testing, with no manual feature engineering.

Experimental results

Research questions

• RQ1 Can a two-stage CNN framework achieve accurate and robust automatic segmentation of the left ventricle in CCTA scans?
• RQ2 How does the performance of the proposed CNN-based method compare to manual segmentation by expert observers?
• RQ3 Can the method effectively segment the LV in apical slices and exclude complex structures like papillary muscles and trabeculae carneae?
• RQ4 Does the use of multi-planar patch inputs improve voxel classification accuracy compared to single-plane analysis?

Key findings

• The proposed method achieved a mean Dice coefficient of 0.85 across five test scans, indicating strong overlap with manual reference annotations.
• The mean absolute surface distance (MAD) between automatic and manual segmentation was 1.1 mm, reflecting high geometric accuracy.
• The second observer’s manual segmentation achieved a higher Dice coefficient (0.90) and lower MAD (0.6 mm), indicating room for improvement over the current method.
• The method successfully excluded papillary muscles and trabeculae carneae despite similar intensity to the LV, due to the spatial context provided by multi-planar patches.
• Performance was consistent across all slices, including the apical region, which is often challenging for segmentation methods.
• The method demonstrated comparable performance to prior state-of-the-art methods, such as those using atlas-based or active contour approaches, with similar or better surface distance metrics.