Landmark Detection in Cardiac MRI by Using a Convolutional Neural Network.

Area of Science:

• Cardiovascular Imaging
• Artificial Intelligence in Medicine
• Medical Image Analysis

Background:

• Accurate landmark detection is crucial for quantitative analysis in cardiac magnetic resonance imaging (CMR).
• Manual landmark identification can be time-consuming and subject to inter-observer variability.
• Deep learning offers potential for automating complex image analysis tasks.

Purpose of the Study:

• To develop and validate a convolutional neural network (CNN) for automated landmark detection in CMR.
• To assess the performance of the CNN across various CMR sequences including cine, late gadolinium enhancement (LGE), and T1 mapping.
• To compare the CNN's accuracy against manual landmark assignments by experienced readers.

Main Methods:

• A retrospective dataset of 2329 patients (34,089 images) from two hospitals was used for training.
• A hold-out test set of 531 patients (7723 images) was used for validation.
• CNN models were trained to detect specific landmarks on long-axis and short-axis CMR images, including mitral valve plane points, apical points, RV insertion points, and LV center points. Model outputs were compared to manual labels.

Main Results:

• High detection rates for cardiac landmarks were achieved, ranging from 96.6% to 100% across different views and sequences.
• Euclidean distances between model-derived and manual labels were small (2-3.5 mm), indicating close agreement.
• The CNN's assessment of anterior RV insertion angle and LV length showed no significant difference compared to reader assessments. Inference times were rapid (610 ms with GPU).

Conclusions:

• A CNN effectively performs landmark detection on diverse CMR image types (cine, LGE, T1 mapping).
• The developed CNN demonstrates accuracy comparable to inter-reader variability in CMR landmark identification.
• This automated approach has the potential to enhance efficiency and consistency in cardiac image analysis.