研究目的
To develop a novel machine-learning-based method for automatic segmentation of left ventricular myocardium and measurement of its volume in gated myocardial perfusion SPECT imaging, aiming to improve accuracy and efficiency without manual intervention.
研究成果
The proposed machine-learning-based method achieves high accuracy in segmenting left ventricular myocardium and measuring its volume in gated MPS imaging, with DSC >0.9 and Hausdorff distance <1 cm. It demonstrates feasibility for clinical use by providing automated, efficient, and reproducible quantification without manual intervention, though further validation with larger datasets and diverse pathologies is recommended.
研究不足
The study relies on manual contours from physicians as ground truth, which may have systematic errors and variability. The dataset size is intermediate (56 patients), and future work should include larger, more diverse populations. The method's performance is dependent on the quality of training data, and clinical impact on disease detection needs further investigation.
1：Experimental Design and Method Selection:
The study uses an end-to-end fully convolutional neural network (3D V-Net) for segmentation, with a compound loss function combining binary cross entropy and Dice loss to handle class imbalance and improve performance. The method involves training on clinical contours as ground truth and testing on new SPECT images.
2：Sample Selection and Data Sources:
Retrospective dataset of 32 normal patients (mean age 63 ± 10, 23 males, 9 females) and 24 abnormal patients (mean age 57 ± 10, 17 males, 7 females) with clinically acquired MPS images. Images were acquired using a dual-headed camera (CardioMD, Philips Medical Systems) with standard resting protocol.
3：List of Experimental Equipment and Materials:
SPECT images acquired with a dual-headed camera (CardioMD, Philips Medical Systems), reconstructed using ordered subsets expectation maximization with Butterworth filter. Computational resources include NVIDIA TITAN XP GPU for model training and segmentation.
4：Experimental Procedures and Operational Workflow:
SPECT images are automatically cropped to 32x32x16 voxels to reduce background. The V-Net is trained with Adam optimizer, batch size 20, and 180 epochs. Segmentation involves inputting cropped images to the trained network, thresholding probability maps at 0.5 to generate contours.
5：5 to generate contours.
Data Analysis Methods:
5. Data Analysis Methods: Performance evaluated using Dice similarity coefficient (DSC) and Hausdorff distance for contour accuracy, Pearson correlation for volume measurement, and Bland-Altman plots for error analysis. Leave-one-out cross-validation is used for evaluation.
独家科研数据包，助您复现前沿成果，加速创新突破
获取完整内容
