Abstract
Myocardial contrast echocardiography (MCE) is an imaging technique that assesses left ventricle function and myocardial perfusion for the detection of coronary artery diseases. Automatic MCE perfusion quantification is challenging and requires accurate segmentation of the myocardium from noisy and time-varying images. Random forests (RF) have been successfully applied to many medical image segmentation tasks. However, the pixel-wise RF classifier ignores contextual relationships between label outputs of individual pixels. RF which only utilizes local appearance features is also susceptible to data suffering from large intensity variations. In this paper, we demonstrate how to overcome the above limitations of classic RF by presenting a fully automatic segmentation pipeline for myocardial segmentation in full-cycle 2D MCE data. Specifically, a statistical shape model is used to provide shape prior information that guide the RF segmentation in two ways. First, a novel shape model (SM) feature is incorporated into the RF framework to generate a more accurate RF probability map. Second, the shape model is fitted to the RF probability map to refine and constrain the final segmentation to plausible myocardial shapes. We further improve the performance by introducing a bounding box detection algorithm as a preprocessing step in the segmentation pipeline. Our approach on 2D image is further extended to 2D+t sequence which ensures temporal consistency in the resultant sequence segmentations. When evaluated on clinical MCE data, our proposed method achieves notable improvement in segmentation accuracy and outperforms other state-of-the-art methods including the classic RF and its variants, active shape model and image registration.
Introduction
Myocardial Contrast Echocardiography (MCE) is a vital clinical tool for detecting coronary artery disease. By utilizing microbubbles to visualize blood flow (perfusion) within the heart muscle, it provides critical insights into cardiac health. However, the automated analysis of MCE images is notoriously difficult due to high speckle noise, signal attenuation, and "shadowing" artifacts. Furthermore, anatomical structures like papillary muscles often share similar intensity profiles with the myocardium, leading standard pixel-wise classifiers to produce physically impossible or "leaky" segmentations.
While Random Forests (RF) are effective for local pixel classification, they lack an inherent understanding of global geometry. Manual segmentation remains the clinical gold standard but is time-consuming and subjective. This paper presents a solution that integrates global shape priors into the RF framework to achieve robust, fully automatic segmentation.
Methodology
The proposed pipeline transitions from coarse localization to fine, shape-constrained segmentation through four primary stages:
- CNN-based Bounding Box Detection: A Convolutional Neural Network (based on CaffeNet) automatically locates the heart within the ultrasound frame. This rigid alignment step standardizes pose, scale, and rotation, simplifying the task for subsequent classifiers.
- Random Forest with Shape Model (SM) Features: Unlike standard RFs that only look at local intensities, this method introduces a novel SM feature. For each pixel, the feature calculates the signed distance to a plausible myocardial shape sampled from a statistical model, forcing the RF to learn global anatomical constraints during local classification.
- Statistical Shape Model (SSM) Fitting: The RF output is a probability map. To ensure anatomical validity, a PCA-based shape model is fitted to this map, acting as a filter that ignores outliers and fills gaps caused by weak ultrasound signals.
- Temporal Consistency (2D+t): For video sequences, a temporal constraint is added to the optimization process, ensuring smooth transitions between frames and preventing jitter.
Experimental Setup
The study utilized clinical MCE data from 21 subjects (10 healthy, 11 with coronary artery disease). Evaluation was conducted on two datasets:
- Dataset 1: 242 static 2D images across various apical views and cardiac phases (end-systole and end-diastole).
- Dataset 2: 12 full-cycle 2D+t sequences (averaging 22 frames each). All data were manually annotated by an expert to establish a ground truth for performance metrics.
Results and Discussion
The proposed method (denoted as N3 in comparative tests) demonstrated significant improvements over classic RF, Active Shape Models (ASM), and image registration techniques. Quantitative analysis using the Hausdorff Distance (HD) showed that the integration of shape model features and bounding box detection (N3) resulted in the lowest error rates and the fewest extreme outliers.
Key findings include:
- Feature-Level Integration: Incorporating shape information as a feature within the RF allows the classifier to be aware of global geometry during the initial decision-making process, effectively solving issues like misclassifying papillary muscles.
- Clinical Correlation: Automated measurements of endocardial and myocardial areas showed high correlation with expert manual measurements ($r=0.93$ and $r=0.69$, respectively).
- Efficiency: The pipeline processes an image in approximately 9.2 seconds. While not yet real-time, this speed is sufficient for near-instantaneous clinical review following a scan.
Conclusion
This research successfully demonstrates that marrying Random Forests with statistical shape priors can overcome the inherent noise and artifacts of contrast echocardiography. By constraining the machine learning model with anatomical knowledge, the pipeline achieves state-of-the-art accuracy in a fully automatic fashion. Future work will focus on expanding the dataset for better generalization and extending the framework to 3D (volumetric) ultrasound data, which would provide a more comprehensive view of myocardial perfusion.
Summary updated