[Paper Review] CNN-based Landmark Detection in Cardiac CTA Scans

Julia M. H. Noothout, Bob D. de Vos+3 more|arXiv (Cornell University)|Apr 13, 2018

Medical Imaging and Analysis18 references22 citations

TL;DR

This paper proposes a patch-based fully convolutional neural network (FCNN) that jointly performs regression and classification to detect six clinically relevant anatomical landmarks in cardiac CTA scans. By using classification to identify patches near landmarks and averaging displacement vectors from these patches, the method achieves sub-3mm average localization error, demonstrating high accuracy without preprocessing or prior segmentation.

ABSTRACT

Fast and accurate anatomical landmark detection can benefit many medical image analysis methods. Here, we propose a method to automatically detect anatomical landmarks in medical images. Automatic landmark detection is performed with a patch-based fully convolutional neural network (FCNN) that combines regression and classification. For any given image patch, regression is used to predict the 3D displacement vector from the image patch to the landmark. Simultaneously, classification is used to identify patches that contain the landmark. Under the assumption that patches close to a landmark can determine the landmark location more precisely than patches farther from it, only those patches that contain the landmark according to classification are used to determine the landmark location. The landmark location is obtained by calculating the average landmark location using the computed 3D displacement vectors. The method is evaluated using detection of six clinically relevant landmarks in coronary CT angiography (CCTA) scans: the right and left ostium, the bifurcation of the left main coronary artery (LM) into the left anterior descending and the left circumflex artery, and the origin of the right, non-coronary, and left aortic valve commissure. The proposed method achieved an average Euclidean distance error of 2.19 mm and 2.88 mm for the right and left ostium respectively, 3.78 mm for the bifurcation of the LM, and 1.82 mm, 2.10 mm and 1.89 mm for the origin of the right, non-coronary, and left aortic valve commissure respectively, demonstrating accurate performance. The proposed combination of regression and classification can be used to accurately detect landmarks in CCTA scans.

Motivation & Objective

• To develop an accurate, end-to-end method for automatic anatomical landmark detection in cardiac CTA scans.
• To overcome limitations of conventional machine learning methods that rely on handcrafted features and exhaustive search.
• To eliminate the need for preprocessing steps such as vessel segmentation or initial landmark localization.
• To improve localization accuracy by combining classification (to identify relevant patches) with regression (to predict 3D displacement).
• To evaluate the method on six clinically relevant landmarks in coronary arteries and the aortic root.

Proposed method

• A fully convolutional neural network (FCNN) is trained to predict a 3D displacement vector from the center of a 3D image patch to the target landmark.
• Simultaneously, the network performs binary classification to determine whether a given patch contains the landmark.
• Only patches classified as positive (containing the landmark) are used to compute the final landmark location via averaging their predicted displacement vectors.
• The network is trained using a combination of regression loss for displacement prediction and classification loss for patch relevance.
• A log-transform is applied to large displacement vectors during training to stabilize learning and reduce the impact of outliers.
• The method is evaluated on 3D CTA volumes at three isotropic resolutions: 1 mm, 1.5 mm, and 3 mm.

Experimental results

Research questions

• RQ1 Can a joint regression and classification FCNN approach improve landmark detection accuracy in cardiac CTA compared to standalone regression or classification?
• RQ2 Does using only patches classified as near the landmark lead to better localization than using all patches?
• RQ3 How does image resolution affect the performance of the proposed landmark detection method?
• RQ4 Can the method achieve high accuracy without requiring preprocessing steps like vessel segmentation or initial landmark seeding?
• RQ5 Is the log-transform of displacement vectors effective in improving training stability and detection accuracy?

Key findings

• The method achieved an average Euclidean distance error of 2.19 mm for the right coronary ostium and 2.88 mm for the left coronary ostium.
• The bifurcation of the left main coronary artery was localized with a mean error of 3.78 mm.
• The origins of the right, non-coronary, and left aortic valve commissures were detected with errors of 1.82 mm, 2.10 mm, and 1.89 mm, respectively.
• Images resized to 1.5 mm isotropic resolution yielded more accurate results than those at 3 mm resolution, while 1 mm resolution did not provide sufficient context for optimal performance.
• The method outperformed or matched prior state-of-the-art methods without requiring preprocessing such as segmentation or initial landmark localization.
• The combination of classification and regression significantly improved accuracy, as classification helped filter out low-quality regression predictions from distant patches.