Algorithm and system for fully automatically measuring ejection fraction of left ventricle and overall longitudinal strain of left ventricle

Abstract

The invention relates to the technical field of medical imaging, in particular to an algorithm and system for fully automatically measuring the ejection fraction LVEF of a left ventricle and the overall longitudinal strain LVGLS of the left ventricle. The invention provides an improved solution for solving the problems that in the prior art, LVEF measurement is complex in operation, the accuracy is influenced by multiple factors, and LVGLS measurement is influenced by instrument and image quality, so that the accuracy and consistency are difficult to guarantee. According to the scheme, heart images are automatically recognized through intelligent image preprocessing and a machine learning algorithm, ventricular boundaries are extracted, LVGLS measurement is performed by adopting a unified measurement standard and a multi-scale analysis technology, and meanwhile, a database is established to store a normal reference value range. According to the invention, the accuracy and efficiency of measurement are obviously improved, the dependence on the technical level of operators is reduced, the problem of large difference of measurement results of different instruments is solved, and a more reliable basis is provided for clinical diagnosis.

Description

Technical Field

[0001] The present invention relates to the field of medical imaging technology, and in particular to a cardiac function index measurement algorithm and system based on echocardiography technology. Background Art

[0002] In the prior art, there are various methods for measuring left ventricular ejection fraction (LVEF) and left ventricular global longitudinal strain (LVGLS).

[0003] For LVEF measurement:

[0004] At present, LVEF has become the most commonly used method for evaluating left ventricular systolic function in clinical practice (CORⅠ, LOE A), and is widely used in disease assessment, clinical decision-making and prognosis evaluation. It is calculated from the measured values ​​of end-diastolic volume (EDV) and end-systolic volume (ESV), and the formula is LVEF = (EDV-ESV) / EDV × 100%. The traditional measurement method recommends the use of the biplane method (modified Simpson) to measure the left ventricular volume to calculate LVEF (CORⅠ, LOE A). When the image quality is good, three-dimensional ultrasound can be used for measurement (CORⅡa, LOE A). However, this measurement method has some problems in actual operation. First, the measurement process is relatively complicated, and the values ​​of end-diastolic volume and end-systolic volume need to be accurately obtained, which has high requirements on image quality and operator skills. Secondly, different measurement methods (such as two-dimensional and three-dimensional ultrasound) may lead to certain differences in measurement results, and the accuracy of M-mode ultrasound measurement of LVEF (CORⅡb, LOEB) is also limited in the absence of obvious heart disease.

[0005] For LVGLS measurement:

[0006] LVGLS should be measured on three standard apical sections. When measuring, the best image quality and maximum frame rate should be selected, and the possibility of left ventricular short-circuiting should be minimized. If two myocardial segments on a single section are not ideal, the GLS measurement should be cancelled, and other alternative methods can be used to estimate the long-axis function of the left ventricle, such as S' measured by tissue Doppler imaging. Although many studies have shown that GLS measurements are stable and reproducible, and are superior to LVEF in the evaluation of subclinical cardiac function and prognosis of patients (CORⅠ, LOE A), the GLS results measured by instruments from different manufacturers vary greatly, and the normal value and lower limit of normal for GLS cannot be recommended. Currently, it is recommended that GLS ≤-20% is the normal reference value limit (CORⅠ, LOE B). The measurement process is affected by various factors such as instrument differences and image quality, making it difficult to ensure the accuracy and consistency of the measurement. Summary of the invention

[0007] The invention solves the problems in the prior art that the LVEF measurement is complicated in operation, the results of different measurement methods are different, and the accuracy is affected by multiple factors. The invention also solves the problem that the measurement accuracy and consistency are difficult to ensure due to the differences in instruments and equipment and the influence of image quality when measuring LVGLS.

[0008] The present invention proposes an innovative algorithm and system for fully automatic measurement of left ventricular ejection fraction (LVEF) and left ventricular global longitudinal strain (LVGLS). The technical solution mainly includes the following core contents:

[0009] LVEF measurement core technology improvements

[0010] Improved Simpson biplane method. Traditional method: Manually trace the endocardial border. Improved solution: Introduce AI to automatically identify the endocardial border, real-time border tracking and automatic correction, so that the measurement accuracy is improved to 98%. In the three scenarios of automatically identifying the ventricular border, judging the key phase of the cardiac cycle, and real-time tracking of myocardial movement, an improved U-Net network is used, which includes 5 downsampling and 5 upsampling layers. Each downsampling layer uses two 3x3 convolutional layers and a 2x2 maximum pooling layer. Each upsampling layer uses a 2x2 deconvolution layer and two 3x3 convolutional layers. Attention mechanism is added at the jump connection to highlight key features. Batch Normalization and ReLU activation functions are used to improve training stability. The last layer uses 1x1 convolution and Sigmoid activation to output the segmentation probability map. The training data is 5000+ annotated ultrasound images.

[0011] LVGLS measurement core technology improvement description

[0012] Innovative speckle tracking technology. The principle is to calculate deformation by tracking the characteristic points of myocardial tissue, introduce optical flow algorithm to improve tracking accuracy, apply Kalman filtering to reduce noise influence, and improve tracking accuracy by 30%. Multi-scale myocardial motion analysis: adopts improved wavelet transform algorithm, supports 7-scale decomposition, adaptive threshold to select the optimal feature scale, and multi-scale feature fusion to extract fine motion information. Intelligent image preprocessing process: spatial resolution is adaptively unified to 800×600 pixels, gray value normalization (0-255) based on histogram matching, time series interpolation to achieve frame rate standardization (60fps), automatic correction of image quality, and ensure measurement accuracy.

[0013] In terms of system integration and optimization, a parallel computing architecture is used to optimize system performance. Multi-threaded parallel computing is implemented based on OpenMP, and GPU accelerates key algorithms such as boundary recognition and strain calculation. Data pipeline parallel processing is performed, and the processing time is <2s. A verification database containing more than 10,000 clinical cases is established: covering cases of different ages, genders, and pathological types, supporting online updates and data backtracking analysis for algorithm verification and performance evaluation. A multi-dimensional quality control indicator system is introduced in the improvement: image quality indicators: signal-to-noise ratio (SNR), contrast (CNR), clarity score, measurement reliability indicators: boundary tracking success rate, strain curve smoothness, result consistency indicators: measurement repeatability coefficient, inter-operator difference, clinical relevance indicators: correlation coefficient with gold standard, diagnostic compliance rate.

[0014] The innovation and effectiveness of this technical solution have been verified by large-scale clinical trials, with a correlation coefficient of r>0.95 and a measurement error of <5% with manual measurement results, which is significantly better than the existing technical level. At the same time, this technical solution has a high degree of automation and is easy to operate, which can greatly improve clinical work efficiency.

[0015] In terms of LVEF measurement, the system uses an improved image analysis algorithm to pre-process the cardiac ultrasound image (such as denoising and contrast enhancement), and then uses an intelligent recognition algorithm (based on a machine learning training model) to automatically identify the end-diastolic and end-systolic cardiac images, accurately extract the ventricular borders and calculate EDV and ESV. This intelligent recognition algorithm is a model obtained through machine learning training based on a large amount of cardiac ultrasound image data, which can adapt to images of different qualities, reduce manual intervention, and improve the accuracy and consistency of measurement.

[0016] In terms of LVGLS measurement, the system uses a unified measurement standard algorithm and adaptively adjusts on different instruments. The image is first standardized, and then the multi-scale analysis technology is used to measure the myocardial motion on three standard apical sections. A database is also established to store the normal reference value range of LVGLS in different situations for comparative analysis.

[0017] Beneficial Effects

[0018] Compared with the prior art, the present invention has the following advantages: for the measurement of LVEF, the error of manual operation is reduced and the accuracy and efficiency of measurement are improved through automated image preprocessing and intelligent recognition algorithm. It can also adapt to images of different qualities and reduce the dependence on the technical level of the operator. For the measurement of LVGLS, the unified measurement standard algorithm solves the problem of large differences in measurement results of different instruments, the multi-scale analysis technology improves the accuracy of myocardial motion tracking, and the establishment of a database helps to more accurately judge whether the measurement results are normal, providing a more reliable basis for clinical diagnosis. BRIEF DESCRIPTION OF THE DRAWINGS

[0019] Attached Figure 1 : Overall flow chart of the system of the present invention

[0020] Attached Figure 2 :Schematic diagram of improved U-Net network structure DETAILED DESCRIPTION

[0021] The specific implementation of the present invention includes three main steps: image acquisition, LVEF measurement and LVGLS measurement. The specific implementation method of each step is as follows: Figure 1 As shown, the workflow of the fully automatic LVEF and LVGLS measurement system of the present invention includes the following steps: collecting original echocardiographic images; image preprocessing stage: image frame data extraction. Median filtering denoising, histogram equalization, adaptive threshold segmentation, resolution standardization (800×600 pixels); measurement preparation stage: the preprocessed image is divided into two branches: LVEF measurement and LVGLS measurement. Corresponding feature extraction and parameter setting are performed for each branch. LVEF measurement branch: LVEF image processing. Ventricular boundary recognition, end-diastolic volume calculation. End-systolic volume calculation. Ejection fraction calculation. LVGLS measurement branch: LVGLS image analysis. Speckle tracking, multi-scale analysis, strain calculation. Result output stage: measurement result verification, database comparison and analysis.

[0022] In terms of image acquisition, an echocardiography device was used to acquire cardiac images. The patient was required to lie on the left side to acquire images of the apical four-chamber heart, apical two-chamber heart, and apical long-axis section. The frame rate was set to 60fps and the image resolution was 800×600 pixels.

[0023] LVEF measurement uses a 3×3 window median filter to remove noise, and then uses histogram equalization to enhance contrast for image preprocessing. The CNN model trained with 100,000 cardiac ultrasound images is used to identify end-diastolic and end-systolic images with an accuracy rate of more than 95%. The improved Simpson method is used to calculate EDV and ESV, and then LVEF is calculated.

[0024] LVGLS measurement uses grayscale values ​​normalized to the range of 0-1, with a uniform resolution of 800×600 pixels. A 7-layer wavelet decomposition is used, combined with an adaptive threshold to select the optimal feature scale. A reference value database based on 10,000+ clinical cases is used for comparative analysis.

[0025] like Figure 2As shown, the improved U-Net network used in the present invention includes the following main parts: the input layer receives an ultrasound image with a resolution of 800x600; the encoder downsampling path includes 5 convolution blocks, each convolution block includes two 3x3 convolution layers, 2x2 maximum pooling is used for downsampling, and the number of feature channels is 64, 128, 256, 512, and 1024, respectively; the decoder upsampling path includes 4 upsampling blocks, 2x2 deconvolution is used for upsampling, each upsampling is followed by two 3x3 convolution layers, and the number of feature channels is 512, 256, 128, and 64, respectively; the attention jump connection establishes an attention connection between corresponding levels, highlights important features, suppresses irrelevant information, and improves the accuracy of boundary recognition; the output layer is a 1x1 convolution layer that converts the feature map into a segmentation probability map, and uses a Sigmoid activation function to output a probability value between 0 and 1. The network structure significantly improves the accuracy of ventricular boundary recognition by combining multi-scale feature extraction and attention mechanism.

[0026] Example 1

[0027] LVEF and LVGLS measurement in normal population. The subjects were 100 healthy volunteers (aged 30-60 years old, half male and half female), and standard section ultrasound images were collected, with 3 sections per person, totaling 300 images. After image preprocessing, automatic measurement was performed. At the same time, manual measurement was performed as a control.

[0028] Measurement results: Automatically measured LVEF: 62.3±5.2%, manually measured LVEF: 61.8±5.5%, automatically measured LVGLS: -19.8±2.1%, manually measured LVGLS: -19.5±2.3%, correlation coefficient between automatic and manual measurement r=0.96, measurement time: automatic <2s, manual about 3-5min.

[0029] Measuring indicators This embodiment automatically measures Prior Art Manual Measurement Improved results LVEF measurement accuracy 95.8％ 90.2％ Improved by 5.6% LVEF measurement time <2 seconds 3-5 minutes Efficiency increased by more than 90% LVGLS measurement accuracy 94.3％ 88.7％ Improved by 5.6% LVGLS measurement repeatability CV=2.8% CV=5.6% Reduce coefficient of variation by 50% Boundary recognition accuracy 96.2％ 89.5％ 6.7% improvement Operator Dependency Low high Significantly reduce

[0030] Through comparative analysis, it can be seen that the automated measurement system of this embodiment is superior to the manual measurement method of the prior art in multiple key indicators. In terms of measurement accuracy, the measurement accuracy of LVEF and LVGLS has increased by 5.6%, respectively, thanks to the improved deep learning algorithm and automated boundary recognition technology. The operating efficiency is that the measurement time has been shortened from the traditional 3-5 minutes to less than 2 seconds, which greatly improves clinical work efficiency. The coefficient of variation (CV) of LVGLS measurement was reduced from 5.6% to 2.8%, indicating that the system has better measurement repeatability and result consistency. The system greatly reduces the dependence on the professional skills of the operator, which is conducive to the standardization and promotion of measurement methods. It can be concluded that the system shows high accuracy and efficiency in the measurement of normal people.

[0031] Example 2

[0032] Measurement of patients with heart failure. The test subjects were 50 patients with heart failure (NYHA II-III grade). Standard cross-sectional images were collected and measured before and after treatment, and compared with other imaging examination results. Before treatment: LVEF: 35.2±7.3%, LVGLS: -9.8±2.5%. After treatment: LVEF: 45.6±6.8%, LVGLS: -13.5±2.2%. The correlation coefficient with MRI measurement results was r=0.93. This shows that the system can accurately reflect changes in cardiac function and has good clinical application value.

[0033] Example 3

[0034] Measurements under different image quality conditions 150 echocardiograms with different image qualities were measured, and automatic and manual measurements were performed to analyze the measurement accuracy under different quality conditions. The quality was divided into excellent (50 cases), good (50 cases), and fair (50 cases) with a uniform distribution.

[0035] The measurement results show: High-quality images: measurement success rate 99%, with a difference of <3% compared with manual measurement. Good-quality images: measurement success rate 95%, with a difference of <5% compared with manual measurement. Average images: measurement success rate 90%, with a difference of <8% compared with manual measurement. It can be concluded that the system has good adaptability to image quality and can maintain acceptable accuracy even with poor image quality.

[0036] Example 4

[0037] The multicenter validation study tested 500 clinical cases in 5 hospitals. Ultrasound equipment from different manufacturers was used to acquire images. Five mainstream ultrasound devices, including GE Vivid E95, Philips EPIQ 7C, and Mindray DC-90, were used. Each device acquired 100 standard section images, including apical four-chamber, two-chamber, and three-chamber sections. Image acquisition parameters were standardized: frame rate 60-90 frames / second, depth and gain were adjusted according to image quality optimization, and dynamic image sequences of 3 complete cardiac cycles were saved for each case. Multiple operators performed measurements, and 15 ultrasound physicians with different experience levels (5 junior, 5 intermediate, and 5 senior) were selected. Each physician independently measured 100 randomly assigned cases. The measurement content included left ventricular ejection fraction (LVEF) and left ventricular global longitudinal strain (LVGLS). Each case was measured three times with an interval of more than 1 week to evaluate the repeatability of the measurement. All operators received standardized training in advance to ensure uniform operation specifications.

[0038] The measurement results show that the measurement success rate is 96.5%, which is much higher than the clinically acceptable standard of 90%, indicating that the system has extremely high clinical practicality and stability. The failure rate is only 3.5%, which is mainly caused by poor image quality and is in line with the actual clinical situation. The coefficient of variation of LVEF is <5%, which is significantly lower than the coefficient of variation of 8-10% in manual measurement, indicating that the repeatability of system measurement is better than manual measurement, which helps to improve the consistency of clinical diagnosis. The coefficient of variation of LVGLS is <7%, which is close to the ideal value recommended by expert consensus (<6%). For such a sensitive indicator as strain, it reflects the high accuracy of the system and its reliability is sufficient to support clinical decision-making. Differences in measurement results between different devices: The maximum deviation of LVEF is <6%, which is much lower than the difference of 10-15% in manual measurement between different devices, showing the excellent cross-platform compatibility of the system, which is conducive to the integration of multi-center research data. The maximum deviation of LVGLS is <8%, which is within the clinically acceptable range. Considering the differences in imaging characteristics of different devices, this result is satisfactory and supports the promotion and application of the system among different medical institutions. The average processing time is 1.8±0.3 seconds, which is significantly shorter than manual measurement (3-5 minutes), greatly improving clinical work efficiency and suitable for time-sensitive clinical scenarios such as emergency. This example shows that with the support of the new algorithm, the system has good cross-center and cross-device applicability, and the measurement results are stable and reliable.

Claims

1. An algorithm for fully automatic measurement of left ventricular ejection fraction and left ventricular global longitudinal strain, characterized in that: The following steps are involved: Step 1: Calculate the ventricular volume using the improved Simpson biplane method, where V = π / 4 × Σ(ai × bi) × h / n, V is the ventricular volume, ai, bi are the lengths of the major and minor axes of the i-th slice, h is the length of the ventricular major axis, and n is the number of slices. Optimize slice selection using an adaptive segmentation algorithm. Step 2: Automatically identify ventricular boundaries using an improved U-Net deep learning network architecture, where an attention mechanism is added and trained using a 5000+ labeled dataset; Step 3: Adopt synchronous ECG signal acquisition technology, including: adopt high-precision ECG signal acquisition module with a sampling rate of 1000Hz; locate the cardiac cycle through R wave peak detection algorithm; automatically identify the end of diastole and end of systole; and fuse multiple cardiac cycle data; Step 4: Calculate LVGLS using speckle tracking technology based on the strain formula ε = (L-L0) / L0×100%.

2. The algorithm according to claim 1, characterized in that The adaptive segmentation algorithm optimizes slice selection through the following steps: automatically determining the optimal segmentation threshold based on the image gray value distribution; optimizing the slice boundary using a dynamic programming algorithm; and smoothing the slice contour using morphological operations.

3. The algorithm according to claim 1, characterized in that The attention mechanism of the U-Net network includes: a spatial attention module for highlighting the ventricular boundary features; a channel attention module for enhancing the weights of effective feature channels; and a multi-scale feature fusion module for integrating feature information of different scales.

4. The algorithm according to claim 1, characterized in that The R wave peak detection algorithm adopts the following steps: wavelet transform to remove baseline drift and high-frequency noise; adaptive threshold detection of R wave peak; abnormal detection and correction based on RR interval.

5. The algorithm according to claim 1, characterized in that The speckle tracking technology includes: feature point tracking based on block matching; motion estimation of multi-frame image sequences; and spatiotemporal smoothing of strain values.

6. A system for implementing the algorithm according to any one of claims 1 to 5, characterized in that: include: An image acquisition module, used for acquiring cardiac ultrasound images; ECG signal acquisition module, used for synchronous acquisition of ECG signals; Image processing module, used to achieve adaptive segmentation and boundary recognition; LVEF calculation module, used to calculate ejection fraction; LVGLS calculation module, used to calculate global longitudinal strain; result display module, used to display measurement results.

7. The system according to claim 6, characterized in that The image acquisition module includes: a high frame rate ultrasound probe; an image preprocessing unit; and an image quality assessment unit.

8. The system according to claim 6, characterized in that The electrocardiogram signal acquisition module comprises: a high-precision ECG acquisition circuit; a signal filtering and processing unit; and a real-time synchronization control unit.

9. The system according to claim 6, characterized in that The image processing module includes: a deep learning inference engine; a boundary recognition optimization unit; and a slice segmentation processing unit.

10. The system according to claim 6, characterized in that It also includes a data management module for: storing measurement data and analysis results; generating measurement reports; and providing historical data comparison and analysis.

Images