Machine learning algorithm using publicly available echo database for simplified "visual estimation" of left ventricular ejection fraction.

BACKGROUND: Left ventricular ejection fraction calculation automation typically requires complex algorithms and is dependent of optimal visualization and tracing of endocardial borders. This significantly limits usability in bedside clinical applications, where ultrasound automation is needed most. AIM: To create a simple deep learning (DL) regression-type algorithm to visually estimate left ventricular (LV) ejection fraction (EF) from a public database of actual patient echo examinations and compare results to echocardiography laboratory EF calculations. METHODS: A simple DL architecture previously proven to perform well on ultrasound image analysis, VGG16, was utilized as a base architecture running within a long short term memory algorithm for sequential image (video) analysis. After obtaining permission to use the Stanford EchoNet-Dynamic database, researchers randomly removed approximately 15% of the approximately 10036 echo apical 4-chamber videos for later performance testing. All database echo examinations were read as part of comprehensive echocardiography study performance and were coupled with EF, end systolic and diastolic volumes, key frames and coordinates for LV endocardial tracing in csv file. To better reflect point-of-care ultrasound (POCUS) clinical settings and time pressure, the algorithm was trained on echo video correlated with calculated ejection fraction without incorporating additional volume, measurement and coordinate data. Seventy percent of the original data was used for algorithm training and 15% for validation during training. The previously randomly separated 15% (1263 echo videos) was used for algorithm performance testing after training completion. Given the inherent variability of echo EF measurement and field standards for evaluating algorithm accuracy, mean absolute error (MAE) and root mean square error (RMSE) calculations were made on algorithm EF results compared to Echo Lab calculated EF. Bland-Atlman calculation was also performed. MAE for skilled echocardiographers has been established to range from 4% to 5%. RESULTS: The DL algorithm visually estimated EF had a MAE of 8.08% (95%CI 7.60 to 8.55) suggesting good performance compared to highly skill humans. The RMSE was 11.98 and correlation of 0.348. CONCLUSION: This experimental simplified DL algorithm showed promise and proved reasonably accurate at visually estimating LV EF from short real time echo video clips. Less burdensome than complex DL approaches used for EF calculation, such an approach may be more optimal for POCUS settings once improved upon by future research and development.

Deep Learning Pitfall: Impact of Novel Ultrasound Equipment Introduction on Algorithm Performance and the Realities of Domain Adaptation.

OBJECTIVES: To test deep learning (DL) algorithm performance repercussions by introducing novel ultrasound equipment into a clinical setting. METHODS: Researchers introduced prospectively obtained inferior vena cava (IVC) videos from a similar patient population using novel ultrasound equipment to challenge a previously validated DL algorithm (trained on a common point of care ultrasound [POCUS] machine) to assess IVC collapse. Twenty-one new videos were obtained for each novel ultrasound machine. The videos were analyzed for complete collapse by the algorithm and by 2 blinded POCUS experts. Cohen's kappa was calculated for agreement between the 2 POCUS experts and DL algorithm. Previous testing showed substantial agreement between algorithm and experts with Cohen's kappa of 0.78 (95% CI 0.49-1.0) and 0.66 (95% CI 0.31-1.0) on new patient data using, the same ultrasound equipment. RESULTS: Challenged with higher image quality (IQ) POCUS cart ultrasound videos, algorithm performance declined with kappa values of 0.31 (95% CI 0.19-0.81) and 0.39 (95% CI 0.11-0.89), showing fair agreement. Algorithm performance plummeted on a lower IQ, smartphone device with a kappa value of -0.09 (95% CI -0.95 to 0.76) and 0.09 (95% CI -0.65 to 0.82), respectively, showing less agreement than would be expected by chance. Two POCUS experts had near perfect agreement with a kappa value of 0.88 (95% CI 0.64-1.0) regarding IVC collapse. CONCLUSIONS: Performance of this previously validated DL algorithm worsened when faced with ultrasound studies from 2 novel ultrasound machines. Performance was much worse on images from a lower IQ hand-held device than from a superior cart-based device.

Making Artificial Intelligence Lemonade Out of Data Lemons: Adaptation of a Public Apical Echo Database for Creation of a Subxiphoid Visual Estimation Automatic Ejection Fraction Machine Learning Algorithm.

OBJECTIVES: A paucity of point-of-care ultrasound (POCUS) databases limits machine learning (ML). Assess feasibility of training ML algorithms to visually estimate left ventricular ejection fraction (EF) from a subxiphoid (SX) window using only apical 4-chamber (A4C) images. METHODS: Researchers used a long-short-term-memory algorithm for image analysis. Using the Stanford EchoNet-Dynamic database of 10,036 A4C videos with calculated exact EF, researchers tested 3 ML training permeations. First, training on unaltered Stanford A4C videos, then unaltered and 90° clockwise (CW) rotated videos and finally unaltered, 90° rotated and horizontally flipped videos. As a real-world test, we obtained 615 SX videos from Harbor-UCLA (HUCLA) with EF calculations in 5% ranges. Researchers performed 1000 randomizations of EF point estimation within HUCLA EF ranges to compensate for ML and HUCLA EF mismatch, obtaining a mean value for absolute error (MAE) comparison and performed Bland-Altman analyses. RESULTS: The ML algorithm EF mean MAE was estimated at 23.0, with a range of 22.8-23.3 using unaltered A4C video, mean MAE was 16.7, with a range of 16.5-16.9 using unaltered and 90° CW rotated video, mean MAE was 16.6, with a range of 16.3-16.8 using unaltered, 90° CW rotated and horizontally flipped video training. Bland-Altman showed weakest agreement at 40-45% EF. CONCLUSIONS: Researchers successfully adapted unrelated ultrasound window data to train a POCUS ML algorithm with fair MAE using data manipulation to simulate a different ultrasound examination. This may be important for future POCUS algorithm design to help overcome a paucity of POCUS databases.

Development of a Deep Learning Network to Classify Inferior Vena Cava Collapse to Predict Fluid Responsiveness.

OBJECTIVES: To create a deep learning algorithm capable of video classification, using a long short-term memory (LSTM) network, to analyze collapsibility of the inferior vena cava (IVC) to predict fluid responsiveness in critically ill patients. METHODS: We used a data set of IVC ultrasound (US) videos to train the LSTM network. The data set was created from IVC US videos of spontaneously breathing critically ill patients undergoing intravenous fluid resuscitation as part of 2 prior prospective studies. We randomly selected 90% of the IVC videos to train the LSTM network and 10% of the videos to test the LSTM network's ability to predict fluid responsiveness. Fluid responsiveness was defined as a greater than 10% increase in the cardiac index after a 500-mL fluid bolus, as measured by bioreactance. RESULTS: We analyzed 211 videos from 175 critically ill patients: 191 to train the LSTM network and 20 to test it. Using standard data augmentation techniques, we increased our sample size from 191 to 3820 videos. Of the 175 patients, 91 (52%) were fluid responders. The LSTM network was able to predict fluid responsiveness moderately well, with an area under the receiver operating characteristic curve of 0.70 (95% confidence interval [CI], 0.43-1.00), a positive likelihood ratio of infinity, and a negative likelihood ratio of 0.3 (95% CI, 0.12-0.77). In comparison, point-of-care US experts using video review offline and manual diameter measurement via software caliper tools achieved an area under the receiver operating characteristic curve of 0.94 (95% CI, 0.83-0.99). CONCLUSIONS: We demonstrated that an LSTM network can be trained by using videos of IVC US to classify IVC collapse to predict fluid responsiveness. Our LSTM network performed moderately well given the small training cohort but worse than point-of-care US experts. Further training and testing of the LSTM network with a larger data sets is warranted.

Are Convolutional Neural Networks Trained on ImageNet Images Wearing Rose-Colored Glasses?: A Quantitative Comparison of ImageNet, Computed Tomographic, Magnetic Resonance, Chest X-Ray, and Point-of-Care Ultrasound Images for Quality.

OBJECTIVES: Deep learning for medical imaging analysis uses convolutional neural networks pretrained on ImageNet (Stanford Vision Lab, Stanford, CA). Little is known about how such color- and scene-rich standard training images compare quantitatively to medical images. We sought to quantitatively compare ImageNet images to point-of-care ultrasound (POCUS), computed tomographic (CT), magnetic resonance (MR), and chest x-ray (CXR) images. METHODS: Using a quantitative image quality assessment technique (Blind/Referenceless Image Spatial Quality Evaluator), we compared images based on pixel complexity, relationships, variation, and distinguishing features. We compared 5500 ImageNet images to 2700 CXR, 2300 CT, 1800 MR, and 18,000 POCUS images. Image quality results ranged from 0 to 100 (worst). A 1-way analysis of variance was performed, and the standardized mean-difference effect size value (d) was calculated. RESULTS: ImageNet images showed the best image quality rating of 21.7 (95% confidence interval [CI], 0.41) except for CXR at 13.2 (95% CI, 0.28), followed by CT at 35.1 (95% CI, 0.79), MR at 31.6 (95% CI, 0.75), and POCUS at 56.6 (95% CI, 0.21). The differences between ImageNet and all of the medical images were statistically significant (P ≤ .000001). The greatest difference in image quality was between ImageNet and POCUS (d = 2.38). CONCLUSIONS: Point-of-care ultrasound (US) quality is significantly different from that of ImageNet and other medical images. This brings considerable implications for convolutional neural network training with medical images for various applications, which may be even more significant in the case of US images. Ultrasound deep learning developers should consider pretraining networks from scratch on US images, as training techniques used for CT, CXR, and MR images may not apply to US.

Artificial intelligence versus expert: a comparison of rapid visual inferior vena cava collapsibility assessment between POCUS experts and a deep learning algorithm.

OBJECTIVES: We sought to create a deep learning algorithm to determine the degree of inferior vena cava (IVC) collapsibility in critically ill patients to enable novice point-of-care ultrasound (POCUS) providers. METHODS: We used publicly available long short term memory (LSTM) deep learning basic architecture that can track temporal changes and relationships in real-time video, to create an algorithm for ultrasound video analysis. The algorithm was trained on public domain IVC ultrasound videos to improve its ability to recognize changes in varied ultrasound video. A total of 220 IVC videos were used, 10% of the data was randomly used for cross correlation during training. Data were augmented through video rotation and manipulation to multiply effective training data quantity. After training, the algorithm was tested on the 50 new IVC ultrasound video obtained from public domain sources and not part of the data set used in training or cross validation. Fleiss' κ was calculated to compare level of agreement between the 3 POCUS experts and between deep learning algorithm and POCUS experts. RESULTS: There was very substantial agreement between the 3 POCUS experts with κ = 0.65 (95% CI = 0.49-0.81). Agreement between experts and algorithm was moderate with κ = 0.45 (95% CI = 0.33-0.56). CONCLUSIONS: Our algorithm showed good agreement with POCUS experts in visually estimating degree of IVC collapsibility that has been shown in previously published studies to differentiate fluid responsive from fluid unresponsive septic shock patients. Such an algorithm could be adopted to run in real-time on any ultrasound machine with a video output, easing the burden on novice POCUS users by limiting their task to obtaining and maintaining a sagittal proximal IVC view and allowing the artificial intelligence make real-time determinations.

Are All Deep Learning Architectures Alike for Point-of-Care Ultrasound?: Evidence From a Cardiac Image Classification Model Suggests Otherwise.

OBJECTIVES: Little is known about optimal deep learning (DL) approaches for point-of-care ultrasound (POCUS) applications. We compared 6 popular DL architectures for POCUS cardiac image classification to determine whether an optimal DL architecture exists for future DL algorithm development in POCUS. METHODS: We trained 6 convolutional neural networks (CNNs) with a range of complexities and ages (AlexNet, VGG-16, VGG-19, ResNet50, DenseNet201, and Inception-v4). Each CNN was trained by using images of 5 typical POCUS cardiac views. Images were extracted from 225 publicly available deidentified POCUS cardiac videos. A total of 750,018 individual images were extracted, with 90% used for model training and 10% for cross-validation. The training time and accuracy achieved were tracked. A real-world test of the algorithms was performed on a set of 125 completely new cardiac images. Descriptive statistics, Pearson R values, and κ values were calculated for each CNN. RESULTS: Accuracy ranged from 96% to 85.6% correct for the 6 CNNs. VGG-16, one of the oldest and simplest CNNs, performed best at 96% correct with 232 minutes to train (R = 0.97; κ = 0.95; P < .00001). The worst-performing CNN was the newer DenseNet201, with 85.6% accuracy and 429 minutes to train (R = 0.92; κ = 0.82; P < .00001). CONCLUSIONS: Six common image classification DL algorithms showed considerable variability in their accuracy and training time when trained and tested on identical data, suggesting that not all will perform optimally for POCUS DL applications. Contrary to well-established accuracies for CNNs, more modern and deeper algorithms yielded poorer results.