Predictive Quantum Artificial Intelligence Lab 1950.Ai Launches to Advance AI through Research and Collaboration

The 17th International Conference on Hybrid Artificial Intelligence Systems (HAIS 2022), hosted by 1950.ai and led by Dr. Shahid Masood, convened in Salamanca, Spain. Over 200 global experts gathered to discuss AI advancements, ethics, and applications in healthcare, finance, and autonomous systems, solidifying Salamanca's status as a hub for AI research and innovation.

Brussels, Belgium--(Newsfile Corp. - May 19, 2024) - Dr. Shahid Masood, a renowned AI expert has announced the launch of the Predictive Quantum Artificial Intelligence Lab 1950.Ai. The lab is dedicated to advancing the field of predictive AI through research and collaboration.

The lab's mission is to harness the power of AI to make predictions and drive decision-making in various industries, from healthcare to finance to transportation. Through its research reports, the lab aims to provide cutting-edge insights that will help shape the future of this field.

Dr. Shahid Masood explaining quantum computing.

The lab's diverse team of researchers, scientists, academics, and analysts is dedicated to advancing the field of predictive AI. They come from various backgrounds and have expertise in different areas of AI, including machine learning, deep learning, natural language processing, and computer vision.

The lab is located in the heart of Brussels, Belgium where the digital landscape is constantly evolving.

The lab is currently working on several research projects, including the development of predictive models for disease diagnosis, the optimization of transportation networks, and the prediction of stock prices. These projects hope to have a significant impact on various industries and help shape the future of AI.

Dr. Shahid Masood explaining artificial intelligence at a conference.

Dr. Masood is excited about the potential of the lab and its impact on the world. He believes that predictive AI has the power to transform industries and improve people's lives. He hopes that the lab's research will inspire others to join him in this quest to advance AI through research and collaboration.

For more information, please contact:

Webmail: ceo@1950.ai Person Name: Dr. Shahid Masood Website URL: https://www.1950.ai/ Youtube: https://www.youtube.com/@DrShahidMasoodYouTube Company Name: 1950.ai

To view the source version of this press release, please visit https://www.newsfilecorp.com/release/209763

Go here to read the rest:
Predictive Quantum Artificial Intelligence Lab 1950.Ai Launches to Advance AI through Research and Collaboration - InvestorsObserver

Director of National Intelligence Avril Haines testifying before a Senate hearing earlier this month. During a May 15 hearing, she identified Russia as the greatest foreign threat to this year's U.S. elections. Win McNamee/Getty Images hide caption

Director of National Intelligence Avril Haines testifying before a Senate hearing earlier this month. During a May 15 hearing, she identified Russia as the greatest foreign threat to this year's U.S. elections.

U.S. elections face more threats than ever from foreign actors, enabled by rapid developments in artificial intelligence, the country's top intelligence official told lawmakers on Wednesday.

Federal, state and local officials charged with protecting voting integrity face a "diverse and complex" threat landscape, Director of National Intelligence Avril Haines told the Senate Intelligence Committee at a hearing about risks to the 2024 elections. But she also said the federal government "has never been better prepared" to protect elections, thanks to lessons learned since Russia tried to influence voters in 2016.

This year, "Russia remains the most active foreign threat to our elections," Haines said. Using a "vast multimedia influence apparatus" encompassing state media, intelligence services and online trolls, Russia's goals "include eroding trust in U.S. democratic institutions, exacerbating sociopolitical divisions in the United States, and degrading Western support to Ukraine."

But it's a crowded field, with China, Iran and other foreign actors also trying to sway American voters, Haines added.

In addition, she said the rise of new AI technologies that can create realistic "deepfakes" targeting candidates and of commercial firms through which foreign actors can launder their activities are enabling more sophisticated influence operations at larger scale that are harder to attribute.

Wednesday's hearing was the first in a series focused on the election, said committee chair Sen. Mark Warner, D-Va., as lawmakers seek to avoid a repeat of 2016, when Russia's meddling caught lawmakers, officials and social media executives off-guard.

Since then, "the barriers to entry for foreign malign influence have unfortunately become incredibly small," Warner said. Foreign adversaries have more incentives to intervene in U.S. politics in an effort to shape their own national interests, he added, and at the same time, Americans' trust in institutions has eroded across the political spectrum.

Sen. Marco Rubio of Florida, the committee's top Republican, questioned how those tasked with protecting the election would themselves be received in a climate of distrust. He raised the specter of a fake video targeting himself or another candidate in the days before November's election.

"Who is in charge of letting people know, this thing is fake, this thing is not real?" he asked. "And I ask myself, whoever is in charge of it, what are we doing to protect the credibility of the entity that is ... saying it, so that the other side does not come out and say, 'Our own government is interfering in the election'?"

Haines said in some cases it would make sense for her or other federal agencies to debunk false claims, while in others it may be better for state or local officials to speak out.

Read more here:
U.S. elections face more threats from foreign actors and artificial intelligence - NPR

Artificial intelligence is everywhere these days, but the fundamentals of how this influential new technology works can be confusing. Two of the most important fields in AI development are machine learning and its sub-field, deep learning. Heres a quick explanation of what these two important disciplines are, and how theyre contributing to the evolution of automation.

Like It or Not, Your Doctor Will Use AI | AI Unlocked

Its worth reminding ourselves what AI actually is. Proponents of artificial intelligence say they hope to someday create a machine that can think for itself. The human brain is a magnificent instrument, capable of making computations that far outstrip the capacity of any currently existing machine. Software engineers involved in AI development hope to eventually make a machine that can do everything a human can do intellectually but can also surpass it. Currently, the applications of AI in business and government largely amount to predictive algorithms, the kind that suggest your next song on Spotify or try to sell you a similar product to the one you bought on Amazon last week. However, AI evangelists believe that the technology will, eventually, be able to reason and make decisions that are much more complicated. This is where ML and DL come in.

Machine learning (or ML) is a broad category of artificial intelligence that refers to the process by which software programs are taught how to make predictions or decisions. One IBM engineer, Jeff Crume, explains machine learning as a very sophisticated form of statistical analysis. According to Crume, this analysis allows machines to make predictions or decisions based on data. The more information that is fed into the system, the more its able to give us accurate predictions, he says.

Unlike general programming where a machine is engineered to complete a very specific task, machine learning revolves around training an algorithm to identify patterns in data by itself. As previously stated, machine learning encompasses a broad variety of activities.

Deep learning is machine learning. It is one of those previously mentioned sub-categories of machine learning that, like other forms of ML, focuses on teaching AI to think. Unlike some other forms of machine learning, DL seeks to allow algorithms to do much of their work. DL is fueled by mathematical models known as artificial neural networks (ANNs). These networks seek to emulate the processes that naturally occur within the human brainthings like decision-making and pattern identification.

One of the biggest differences between deep learning and other forms of machine learning is the level of supervision that a machine is provided. In less complicated forms of ML, the computer is likely engaged in supervised learninga process whereby a human helps the machine recognize patterns in labeled, structured data, and thereby improve its ability to carry out predictive analysis.

Machine learning relies on huge amounts of training data. Such data is often compiled by humans via data labeling (many of those humans are not paid very well). Through this process, a training dataset is built, which can then be fed into the AI algorithm and used to teach it to identify patterns. For instance, if a company was training an algorithm to recognize a specific brand of car in photos, it would feed the algorithm huge tranches of photos of that car model that had been manually labeled by human staff. A testing dataset is also created to measure the accuracy of the machines predictive powers, once it has been trained.

When it comes to DL, meanwhile, a machine engages in a process called unsupervised learning. Unsupervised learning involves a machine using its neural network to identify patterns in what is called unstructured or raw datawhich is data that hasnt yet been labeled or organized into a database. Companies can use automated algorithms to sift through swaths of unorganized data and thereby avoid large amounts of human labor.

ANNs are made up of what are called nodes. According to MIT, one ANN can have thousands or even millions of nodes. These nodes can be a little bit complicated but the shorthand explanation is that theylike the nodes in the human brainrelay and process information. In a neural network, nodes are arranged in an organized form that is referred to as layers. Thus, deep learning networks involve multiple layers of nodes. Information moves through the network and interacts with its various environs, which contributes to the machines decision-making process when subjected to a human prompt.

Another key concept in ANNs is the weight, which one commentator compares to the synapses in a human brain. Weights, which are just numerical values, are distributed throughout an AIs neural network and help determine the ultimate outcome of that AI systems final output. Weights are informational inputs that help calibrate a neural network so that it can make decisions. MITs deep dive on neural networks explains it thusly:

To each of its incoming connections, a node will assign a number known as a weight. When the network is active, the node receives a different data item a different number over each of its connections and multiplies it by the associated weight. It then adds the resulting products together, yielding a single number. If that number is below a threshold value, the node passes no data to the next layer. If the number exceeds the threshold value, the node fires, which in todays neural nets generally means sending the number the sum of the weighted inputs along all its outgoing connections.

In short: neural networks are structured to help an algorithm come to its own conclusions about data that has been fed to it. Based on its programming, the algorithm can identify helpful connections in large tranches of data, helping humans to draw their own conclusions based on its analysis.

Machine and deep learning help train machines to carry out predictive and interpretive activities that were previously only the domain of humans. This can have a lot of upsides but the obvious downside is that these machines can (and, lets be honest, will) inevitably be used for nefarious, not just helpful, stuffthings like government and private surveillance systems, and the continued automation of military and defense activity. But, theyre also, obviously, useful for consumer suggestions or coding and, at their best, medical and health research. Like any other tool, whether artificial intelligence has a good or bad impact on the world largely depends on who is using it.

See more here:
Machine Learning vs. Deep Learning: What's the Difference? - Gizmodo

For the past year and a half since ChatGPT was released, a scary question has hovered over the heads of major online publishers: What if Google decides to overhaul its core search engine to feature generative artificial intelligence more prominently and breaks our business in the process?

The question speaks to one of the most fragile dependencies in todays online media ecosystem.

Most big publishers, including The New York Times, receive a significant chunk of traffic from people going to Google, searching for something and clicking on articles about it. That traffic, in turn, allows publishers to sell ads and subscriptions, which pay for the next wave of articles, which Google can then show to people who go searching for the next thing.

The whole symbiotic cycle has worked out fine, more or less, for a decade or two. And even when Google announced its first generative A.I. chatbot, Bard, last year, some online media executives consoled themselves with the thought that Google wouldnt possibly put such an erratic and unproven technology into its search engine, or risk mucking up its lucrative search ads business, which generated $175 billion in revenue last year.

But change is coming.

At its annual developer conference on Tuesday, Google announced that it would start showing A.I.-generated answers which it calls A.I. overviews to hundreds of millions of users in the United States this week. More than a billion users will get them by the end of the year, the company said.

The answers, which are powered by Googles Gemini A.I. technology, will appear at the top of the search results page when users search for things like vegetarian meal prep options or day trips in Miami. Theyll give users concise summaries of whatever theyre looking for, along with suggested follow-up questions and a list of links they can click on to learn more. (Users will still get traditional search results, too, but theyll have to scroll farther down the page to see them.)

The addition of these answers is the biggest change that Google has made to its core search results page in years, and one that stems from the companys fixation on shoving generative A.I. into as many of its products as possible. It may also be a popular feature with users Ive been testing A.I. overviews for months through Googles Search Labs program, and have generally found them to be useful and accurate.

We are having trouble retrieving the article content.

Please enable JavaScript in your browser settings.

Thank you for your patience while we verify access. If you are in Reader mode please exit andlog intoyour Times account, orsubscribefor all of The Times.

Thank you for your patience while we verify access.

Already a subscriber?Log in.

Want all of The Times?Subscribe.

More:
Can Google Give A.I. Answers Without Breaking the Web? - The New York Times

Ethics approval

The CMR datasets were acquired retrospectively under the approval of the institutional review boards (IRBs) at each participating institution, including Beijing Fuwai Hospital, Beijing Anzhen Hospital, Guangdong Provincial Peoples Hospital, the 2nd Affiliated Hospital of Harbin Medical University, the First Hospital of Lanzhou University, Renji Hospital, Tongji Hospital and Peking Union Medical College Hospital. Informed consent was waived by the IRBs. Before model training, testing and reader studies, all data underwent deidentification processes.

The CMR database search was performed for all eight centers to identify CVDs and normal controls. All data were anonymized and deidentified, as per the Health Insurance Portability and Accountability Act Safe Harbor provision56. Inclusion criteria were (1) patients with a definitive diagnosis of CVD and (2) patients with CMR scans at baseline before surgical treatment, if any. Exclusion criteria were (1) incomplete cine or LGE modalities, (2) SAX cine with fewer than five views, (3) CMR images with insufficient scan quality, (4) CVD patients missing clinical data and (5) CMR examinations that could not be interpreted and agreed upon by the committee cardiologists according to the diagnostic criteria (Methods). The detailed diagnostic criteria of the 11 types of CVDs and normal controls included in this study was described in Methods. Table 1 and Extended Data Table 1 present the detailed demographics and distribution of the primary dataset and the external validation sets collected from the other seven medical centers across China. To offer a comprehensive perspective on our primary development dataset, we went the extra mile by collecting the LV ejection fraction (LVEF) metric for all 7,900 subjects (including 1,250 normal controls and 6,650 patients with CVD) within the primary dataset. We meticulously summarized the distribution of demographics and LVEF across the 11 specified CVD classes and the normal control class in Supplementary Table 5. Additionally, we generated density plots to illustrate the distribution of LVEF for each class in the primary dataset, offering a more comprehensive representation (Supplementary Fig. 1).

The fresh consecutive testing set is designed to capture the genuine spectrum of disease phenotypes in the real-world clinical prevalence. To offer a thorough understanding of the severity of cases in alignment with real-world clinical prevalence, we have presented five key cardiac function metrics. These metrics include LVEF, LV mass, LVMi (LV mass index), LV end-diastolic volume and LV end-diastolic volume index. Supplementary Table 6 presents the distribution of demographics and the cardiac functions across 11 CVD classes and the normal control class in the fresh consecutive testing set. For improved visualization and clarity, we have depicted the prevalence of the 11 CVD classes in both the fresh consecutive testing set (n=532 patients with CVD) and the primary discovery dataset (n=6,650 patients with CVD) using pie charts in Supplementary Fig. 2. The fresh consecutive testing set offers a representation of the genuine clinical prevalence. Through direct comparison, it is evident that the primary dataset and the consecutive testing set exhibit very similar CVD prevalence and distribution. The top three most prevalent CVDs referred to the CMR examination remain HCM, DCM and CAD.

All images were acquired by breath-holding and electrocardiographic gating. A balanced steady-state free precession sequence was used for cine images with a continuous sampling from the basal to the apical levels on SAX views and two-chamber, three-chamber and 4CH long-axis views. We included cine MRI from two views in this study: the standard SAX cine and the long-axis 4CH cine. The SAX cine clearly depicts the RV and the LV. The 4CH cine shows the four chambers of heart: right atrium, left atrium, RV and LV.

LGE MRI has been established as the gold standard reference for myocardial viability and replacement fibrosis in the myocardium57,58. In our CMR cohorts, the LGE images were obtained using phase-sensitive inversion recovery sequence with a segmented FLASH readout scheme performed 1015min after injection of gadolinium-based contrast with 0.15mmolkg1 per bolus. Gadolinium contrast agents can be used to detect areas of fibrosis, as the prolonged washout of the contrast correlates with a reduction in functional capillary density in the irreversibly injured myocardium59. The SAX LGE used in the study was acquired from the SAX view with the same section thickness, covering the entire left ventricle from the base to the apex (nine parallel views for most cases). Note that LGE is an invasive examination that requires contrast injection and was therefore not performed for normal controls.

The typical CMR scan protocol and scanner parameters for the primary and external validation sets are presented in Supplementary Table 7. Extended Data Fig. 2 shows an illustration of cardiac MRIs (SAX cine, 4CH cine and SAX LGE) utilized in model development. Supplementary Videos 111 demonstrate example CMR of the 11 types of CVDs.

For each patient in the disease cohort, the textual description of the abnormalities in the CMR and the clinical report was extracted as the main reference. Besides that, all CMR records underwent additional annotation procedures. To annotate the disease cohort, a group of certified CMR experts reviewed all records and clinical reports. Every record was randomly assigned to be reviewed by a single physician specifically for this task, not for any other purpose. All annotators received specific instructions and training regarding how to annotate CMR data to improve labeling consistency. The diagnostic criteria we adopted in this study for each CVD class are described in Methods. CMR examinations that could not be interpreted by physicians received further annotation from a consensus committee of board-certified practicing cardiologists (with >15years of experience in CMR reading) working in Fuwai Hospital. The CMR examinations that could not be interpreted or agreed upon by the committee were removed from our dataset.

For the independent gold-standard test dataset with 500 patients (Extended Data Table 6) for humanmachine comparison, six physicians working in the MRI department at Fuwai Hospital contributed directly to its annotation (the six physicians were not involved in dataset annotation as described above). All participating physicians received specific instructions and training regarding how to annotate CMRs to ensure consistency. We divided the physicians into three groups according to their reading experience in CMR: 35years, 510years and more than 10years. CMR physicians in each group reviewed a randomly selected set of the 500 CMRs in a nonrepetitive manner.

The CMR preprocessing pipeline aimed to remove the additional burden of the deep neural network learning to find patterns between images for disease classification. All cardiac MRIs were preprocessed to (1) resample MRI images to the same spatial resolution and (2) localize the heart region of interest (ROI) to a crop image. We detailed the preprocessing step for cine and LGE MRI below and in Extended Data Fig. 4.

SAX cine comprises nine parallel views (for most cases) covering the apical to the basal levels of the LV. Each view contains 25 frames (cardiac phases), leading to 225 images in one single SAX cine record. We examined the representational power of different numbers of input views in developing the classification model. Balancing efficiency and effectiveness, the three-view input scheme achieved a greater representation of SAX cine and therefore is adopted throughout the rest of the study. The three-view input scheme includes the middle layer (the mid slice among the parallel layers spanning from the base to the apex), the second layer above the middle layer and the second layer below the middle layer (Extended Data Fig. 2). We extract the ImagePositionPatient tag and the ImageOrientationPatient tag from each Dicom header to locate the three layers. Then, three-spline interpolation provided by SimpleITK60 library (https://simpleitk.org/) is applied to resample the raw cine MRIs to the same spatial resolution of 0.994mm0.994mm, which is the most common spatial resolution across all subjects investigated in this study. We developed a heart ROI segmentation model (the following section) and used it to localize the region of heart for each cine MRI. The heart ROI segmentations predicted by the AI models were manually checked to ensure their accuracy. The extracted ROIs are padded to keep the aspect ratio the same without distortion, and then resized to 224224. The top and bottom 0.1% of the pixels in cine MRI images are clipped to avoid pixels that are outliners of the distribution. The cine images are scaled between 1 and 255, and then normalized by zero mean and unit variance before feeding them to the model. We sample a clip of 25 frames from each full-length cine sequence using a temporal stride of two, resulting in 13 frames as inputs to model development. The 4CH cine shares the same preprocessing pipeline as SAX cine, except that only one single layer (mid slice) is used to represent the 4CH view. For SAX LGE, all layers covering from the base to the apex of the heart are used for diagnostic model development. The preprocessing steps for SAX LGE are similar to that of cine MRI. We resampled SAX LGE along the z-axis to ensure that each LGE sequence contains nine slices because nine is the most common number of views for SAX LGE included in this study.

We developed heart detection DNN models to automatically extract the heart ROI regions (Extended Data Fig. 4). Three DNN models for SAX cine, 4CH cine and SAX LGE were trained and evaluated, respectively. We applied nnU-Net61 as our model backbone and generated the ground-truth segmentation masks for model supervision using a semi-automatic approach. (1) Automatic localization: for SAX cine and 4CH cine, we selected the pixel region with maximum standard deviation across all frames. These regions localize the heart ROI as heart is a beating organ with high standard deviation in its position. Specifically, for each cine movie sequence (s={{x}_{1},ldots ,{x}_{n}}), we computed a single pixel map of standard deviations across all frames ({x}_{mathrm{std}}=sigma ({{x}_{1},ldots ,{x}_{n}})). This map was used to compute an Otsu threshold to binarize and label regions with the greatest variation in cine modality21. For each cine sequence, a binary segmentation mask of the heart ROI is defined for the length of the cardiac cycle. All segmentation masks went through manual checking. The localization procedure captures the heart ROI in around 90% of cases. The rest of the cases are labeled manually. (2) Manual labeling: we manually drew the bounding box capturing the heart ROI, using 3D Slicer62 and ITK-SNAP63. We used the Scissors tool provided by the Segment Editor in 3D Slicer and the Polygon Inspector in ITK-SNAP to locate heart ROI. A binary segmentation mask was saved for each CMR sequence. For SAX LGE, we manually drew the annotations as model supervision.

In terms of model architecture, the detection model shares the classic U-net64 backbone with three small adjustments: (1) batch normalization is replaced with instance normalization65, (2) rectified linear unit (ReLU) is replaced with leaky ReLU66 as the activation function and (3) additional auxiliary losses are added in the decoder to all but the two lowest resolutions. The model outputs the binary bounding box that extracts the heart ROI. For model training, we adopted Adam optimizer and stochastic gradient descent (SGD) with Nesterov momentum (=0.99). The initial learning rate was set to be 0.01, and the decay of the learning rate followed the Poly learning rate policy67. Batch size was set to 36. Data augmentation included rotations, scaling, gamma correction and mirroring. The loss function is the sum of cross-entropy and Dice loss68.

For models based on cine sequence, we sampled a clip of 13 frames from each 25-frame cine video using a temporal stride of 2 and spatial size of 224224, resulting in 75656 input 3D tokens. The 3D patch partitioning layer obtains tokens, with each patch/token consisting of a 128-dimensional feature. In practice, 3D convolution without overlapping is applied for this tokenization, and the number of output channels is set to be 128 to project the features of each token to a 128-dimension.

The developed model consists of four stages, that is, four video swin transformer blocks. Each stage, besides the last stage, performs 2 spatial downsampling in the patch merging layer. It is worth noting that we do not downsample along the temporal dimension. The patch merging layer concatenates the features of each group of 22 spatially neighboring patches and applies a linear layer to project the concatenated features to half of their dimension. The video swin transformer block consists of a 3D window-based multihead self-attention module and a 3D-shifted window-based multihead self-attention module, followed by a feedforward network, that is, a two-layer multilayer perceptron, with Gaussian error linear unit nonlinearity in between. Layer normalization is applied before each multihead self-attention module and multilayer perceptron, and a residual connection is applied after each module. We used the base version of VST. The number of heads for each stage is 4, 8, 16 and 32. Extended Data Fig. 3a shows the schematic overview of the VST-based framework for modeling SAX cine.

Model performance improved with increasing training data sample size. For the screening model, we used random rotation, random color jitter and adding random number. During each step of SGD in the training process, we perturbed each training sample, cine video sequences, with a random rotation (between 45 and +45 degrees for SAX cine and between 20 to +20 degrees for 4CH cine), random color jitter and with adding a number sampled uniformly between 0.1 and 0.1 to image pixels (pixel values are normalized) to increase or decrease brightness of the images. For LGE, we used random rotation between 45 and +45 degrees, random color jitter and random flip along the z-axis. Data augmentation resulted in improvement for all models.

First, we developed VST-based models for SAX cine, 4CH cine and SAX LGE, respectively. Then, to fuse information from different modalities, we added a global average pooling layer following the last self-attention module for each VST model. This resulted in a 1,024-dimension feature vector from each modality. We further concatenated the 1,024-dimension vectors and added a fully connected layer on top of that to aggregate the features. The final fully connected softmax layer produces a distribution over the output classes. In terms of training, we loaded and froze the pretrained weights of each VST branch from different modalities using transfer learning69 and only finetuned the last fully connected layers for feature aggregation.

Following the classic VST configuration27, we employed an AdamW optimizer using a cosine decay learning rate scheduler and 2.5 epochs of linear warmup. A batch size of 32 was used. The backbone VST is initialized from the ImageNet70 and Kinetics-600 (ref. 71) pretrained model; the head is randomly initialized. Model pretraining plays a strikingly important role in VST-based CMR interpretation. We also found that multiplying the learning rate of the backbone by 0.1 improves performance. Specifically, the initial learning rates for the pretrained backbone and randomly initialized head were set to be 1104 and 1103, respectively. The impact of learning rate modification on the VST backbone was systematically examined as below. We adopt 0.2 stochastic depth rate and 0.05 weight decay for the Swin base model used in this study. To prevent the models from becoming biased toward one class, we balanced the training datasets for both screening and diagnostics using the ClassBalancedDataset sampling strategy72. Each VST branch derived from the single modality was trained for 150 epochs and then fed into the fusion model, following with 20 epochs of finetuning particularly for the fusion layers. For inference, we set the batch size to be one and the number of workers to be four. The training time for model development using four NVIDIA GeForce RTX 3090 graphics processing units with 24GB VRAM was about 77h, and the inference time for each subject was only 0.233s.

The impact of learning rate modification on the VST backbone was systematically examined through a controlled experiment. The experiment encompassed a range of learning rates, from 1102 to 1106, with a focus on their effects on the AI diagnostic model based on SAX cine. The investigation was conducted on the primary cohort (6,650 CVD patients), utilizing a twofold configuration for training and the remaining fold for testing. The model was trained for 150 epochs with five different learning rate initializations for the model backbone: 1102, 1103, 1104 (as applied in this study), 1105 and 1106. Other configurations were kept consistent for a fair and direct comparison, and the training loss for each scheme was plotted for analysis (Supplementary Fig. 3). From the depicted figure, several key observations emerge. When the learning rate is set too high (1102, curve in blue color), the model struggles to converge and the training loss fails to descend, in stark contrast to the more optimal setting of 1104 (curve in green color). Notably, the model under the 1102 learning rate incorrectly classified all samples into the HCM class during testing. Conversely, when the learning rate is set too low (1106, curve in purple color), the loss descends very slowly over the training period. As depicted in the figure, the loss curves for 1105 and 1106 remain at a relatively high level compared with the more effective setting of 1104. Further evaluation included the calculation of F1 and area under the receiver operating characteristic curve scores for the testing fold under the aforementioned experimental settings (Supplementary Fig. 3). Notably, the model trained with a learning rate of 1102 failed to converge and was consequently excluded from the quantitative metrics. According to the evaluation results, the initialized learning rate of 1104 demonstrated superior performance compared with the other settings. Therefore, based on these comprehensive analyses, we selected 1104 as the initialized learning rate for our experiment.

We examined the conventional CNNLSTM architecture in CMR interpretation. The CNNLSTM consists of a DenseNet encoder with 40 layers and a growth rate of 12 for feature extraction and an LSTM for temporal feature aggregation. DenseNet encoder comprised a series of two-dimensional convolutions with kernel sizes 11 and 33 and global average pooling to extract the feature vector for each input frame. For LSTM, the feature vector for each input frame is fed into the LSTM module sequentially. LSTM fuses the feature vectors and produces the final classification score after one fully connected layer. For the training configuration of the CNNLSTM model, we adopt the SGD optimizer with a learning rate of 0.001, a momentum of 0.9 and a weight decay of 0.001. A batch size of four is used for training and one is used for testing. The DenseNet encoder of the CNNLSTM model is initialized from the pretrained model21 and the LSTM component is randomly initialized. We kept data augmentation, the input scheme and computational resources the same as VST models with the only difference: SAX cine inputs are resized to 6464 due to CNNLSTM memory constraints.

The performance of the AI models was evaluated by assessing their sensitivity, specificity, precision and F1 score (harmonic mean of the predictive positive value and sensitivity), with two-sided 95% CIs, as well as the AUC of the ROC with two-sided CIs. The F1 score is complementary to the AUC, which is particularly useful in the setting of multiclass prediction and less sensitive than the AUC in settings of class imbalance. For an aggregate measure of model performance, we computed the class frequency-weighted mean for the F1 score and the AUC73.

The cutoff value was set to 0.5 for screening; the CVD class with the highest probability was the diagnostic prediction. Precision, sensitivity (recall), specificity, PPV, NPV and F1 score of each class are related to true-positive (TP), true-negative (TN), false-positive (FP) and false-negative (FN) rates, with formulas as follows:

$$text{Sensitivity}=frac{mathrm{TP}}{mathrm{TP}+mathrm{FN}},$$

$$text{Specificity}=frac{mathrm{TN}}{mathrm{TN}+mathrm{FP}},$$

$$mathrm{Precision}=frac{mathrm{TP}}{mathrm{TP}+mathrm{FP}},$$

$$mathrm{PPV},=frac{mathrm{TP},}{mathrm{TP}+mathrm{FP}},$$

$$mathrm{NPV},=frac{mathrm{TN},}{mathrm{TN}+mathrm{FN}},$$

$${F}_{1}text{-score}=frac{2times mathrm{Precision}times mathrm{Sensitivity}}{mathrm{Precision}+mathrm{Sensitivity}}.$$

The ROC space is defined by 1specificity and sensitivity as the x axis and the y axis, respectively. It depicts relative trade-offs between true positive and false positive, as the classification threshold goes from zero to one. A random guess will give a point along the diagonal line from the bottom left to the top right. Points above the diagonal line represent good classification results and points below the line represent bad results. We applied the class frequency-weighted F1 score and class frequency-weighted AUC to evaluate the performance of our diagnostic model, with the following formulas:

$${rm{Weighted}},{F}_{1}text{-}{rm{score}}=mathop{sum }limits_{i}^{C}{mathrm{ratio}}_{i}{F}_{1}text{-}{mathrm{score}}_{i},$$

$${rm{Weighted}},mathrm{AUC}=mathop{sum }limits_{i}^{C}{mathrm{ratio}}_{i}{mathrm{AUC}}_{i},$$

where ({{F}_{1}text{-score}}_{i}) and AUCi denote the F1 score and AUC for class i, respectively, and ({mathrm{ratio}}_{i}) denotes a frequency ratio for each class i.

In addition, to improve the model interpretability and visualize the features used by the DNN model that determine the final prediction, we used Grad-CAM29 to localize important regionssaliency regionsby visualizing class-specific gradient information. In Grad-CAM, the neuron importance weight ({alpha }_{k}^{,c}) is estimated as

$${alpha }_{k}^{,c}=frac{1}{Z}sum _{i}sum _{j}frac{partial {y}^{,c}}{partial {A}_{{ij}}^{k}},$$

where yc denotes the gradient score for class (c) before the softmax and Ak denotes the feature map activation of the kth layer. After computing the neuron importance weights for each feature map, we can generate a heat map indicating the significant regions related to class (c) by performing a weighted linear combination of the feature maps, followed with a ReLU activation function as

$${L}_{mathrm{Grad}-mathrm{CAM}}^{c}=mathrm{ReLU}left(sum _{k}{alpha }_{k}^{,c}{A}^{k}right).$$

We then used the Shapley values74 to evaluate the influence of each input modality (SAX cine, 4CH cine and SAX LGE). The Shapley value is a principled attribution method used in AI to quantify the contribution of individual input features by assigning each input modality an importance value for a particular prediction. The definition of the Shapley value75 is given in equations below:

$${{{phi }}}_{i}left(vright)=sum _{Ssubset N{i}}{left(begin{array}{c}n\ 1,left|Sright|,n-left|Sright|-1end{array}right)}^{-1}left(vleft(Scup {i}right)-vleft(Sright)right),$$

where ({phi}_{i}left(vright)) denotes the contribution value of input component i, namely the Shapley value of each input modality (player), (N) is the number of layers and (v) is a function mapping subsets of layers to the real numbers: (v:{2}^{N}to {R}), with (vleft(varnothing right)=0), where (varnothing) denotes the empty set. A set of players is called a coalition. The function (v) is called a characteristic function: if (S) is a coalition of players, then (v(S)), called the worth of coalition (S), describes the total expected sum of payoffs the members of (S) can obtain by cooperation. The sum extends over all subsets (S) of (N) not containing input component i; also note that (left(begin{array}{c}n\ a,{b},{c}end{array}right)) is the multinomial coefficient. This formula can also be interpreted as

$$begin{array}{l}{{{phi }}}_{i}left(vright)=frac{1}{{mathrm{Number}};{rm{of}};{rm{layers}}}\sum _{{mathrm{coalitions}}; {mathrm{including}};i}frac{{mathrm{Marginal}};{mathrm{contribution}}; {mathrm{of}};i;{mathrm{to}};{mathrm{coalition}}}{{mathrm{Number}}; {mathrm{of}}; {mathrm{coalitions}}; {mathrm{excluding}};i; {mathrm{of}}; {mathrm{this}}; {mathrm{size}}}.end{array}$$

The diagnosis of myocardial infarction or ischemic cardiomyopathy is based on the European Society of Cardiology, American College of Cardiology and American Heart Association committee criteria76 with significant stenosis on invasive coronary angiography (CAG) or coronary computed tomography angiography, and CMR showed subendocardial or transmural LGE with matching coronary arteries. We excluded cases without available CAG present or inadequate image quality due to arrhythmia or respiratory motion artifact.

We followed the 2020 American Heart Association and American College of Cardiology guidelines for the diagnosis of patients with HCM77. The clinical diagnosis of HCM was made by CMR showing a maximal end-diastolic wall thickness of 15mm anywhere in the LV, in the absence of another cause of hypertrophy in adults. More limited hypertrophy (1314mm) can be diagnostic when present in family members of a patient with HCM or in conjunction with a positive genetic test.

We excluded cases with the following conditions:

Valvular heart disease (aortic valve stenosis, etc.)

Long-term uncontrolled hypertension

Inflammatory heart disease (sarcoidosis, etc.)

Infiltrative cardiomyopathy (amyloidosis, Fabry disease, etc.)

Septal myectomy or alcohol ablation before CMR

CMR images with poor quality

The diagnosis of DCM is based on the diagnostic criteria of the World Health Organization78. Inclusion criteria were based on enlarged LV end-diastolic dimension (>60mm) and reduced LVEF (<45%). The exclusion criteria were as follows:

Significant stenosis of coronary artery (>50% stenosis, assessed on CAG or coronary computed tomography angiography)

Severe valvular disease, hypertension or congenital heart disease

Evidence of acute or subacute myocarditis (T2 weighted image and laboratory tests)

Any other metabolic disease through medical documentation

Inadequate CMR quality

The diagnosis of LVNC is based on previous studies32,79, as follows:

The presence of noncompacted and compacted LV myocardium with a two-layered appearance, with at least involvement of the LV apex

End-diastolic noncompaction/compaction ratio >2.3 on long-axis views and 3 on SAX views

Noncompacted mass >20% of the global LV mass

No pathologic (pressure/volume load, for example, hypertension) or physiologic (for example, pregnancy and vigorous physical activity) remodeling factors leading to excessive trabeculation

The diagnostic standards for ARVC were based on the revised Task Force Criteria80 score with either two major criteria, one major and two minor criteria or four minor criteria. The major criteria include regional RV akinesia or dyskinesia or dyssynchronous RV contraction, ratio of RV end-diastolic volume to body surface area >110mlm2 (male) or >100mlm2 (female) or RV ejection fraction <40%; fibrous replacement of the RV free wall myocardium, with or without fatty replacement of tissue on endomyocardial biopsy; repolarization abnormalities and depolarization or conduction abnormalities on ECG test.

The diagnosis of CAM is based on endomyocardial biopsy or extracardiac biopsy specimens showing positive birefringence with Congo red staining under polarized light, and with native and enhanced CMR imaging in a pattern consistent with CAM: LV wall thickness of more than 12mm shown by CMR without other known cause, with and without diffuse LGE81.

RCM is characterized by ventricular filling difficulties with increased stiffness of the myocardium. The restrictive cardiomyopathies are defined as restrictive ventricular physiology in the presence of normal or reduced diastolic volumes52,82, as follows:

Nondilated LV or RV with diastolic dysfunction

Bi-atrial dilation

Preserved ejection fraction (LVEF 50%)

We excluded subjects that met the following criteria:

With a reduced LV systolic function

Severe atrial fibrillation

Severe valvular disease, hypertension or congenital heart disease

Significant stenosis of coronary artery.

The diagnosis of PAH is based on the results of right heart catheterization examination. Patients are included in this study if they were clinically diagnosed as PAH83:

Mean pulmonary artery pressure (mPAP) 25mmHg

Pulmonary capillary wedge pressure (PCWP) <15mmHg

Pulmonary vascular resistance (PVR) >3 Wood units at rest

We excluded subjects with the following criteria:

Any evidence of cardiomyopathy, myocarditis, CAD, myocardial infarction, valvular disease, or constrictive pericarditis.

Any evidence of respiratory diseases.

History of cardiac surgery

The diagnosis of Ebsteins anomaly is based on apical displacement of tricuspid valve leaflets (8mmm2) with fibrous and muscular attachments to the underlying myocardium31. Patients with other concomitant malformation (for example, congenitally corrected transposition with Ebsteins anomaly) and history of cardiac surgery were excluded.

The diagnosis of acute myocarditis is based on the diagnostic criteria for clinically suspected myocarditis, as recommended by the European Society of Cardiology Working Group on Myocardial and Pericardial Diseases84, and is fulfilled by meeting the Lake Louise criteria85 or by confirmation through endomyocardial biopsy.

Patients with clinically acute myocarditis had the following: acute chest pain, signs of acute myocardial injury (electrocardiographic changes and/or elevated troponin level) and increased laboratory markers of inflammation (for example, C-reactive protein level). CAD was excluded before cardiac MRI. Patients with preexisting CVD were excluded.

The diagnostic criteria for HHD include (1) a history of prolonged, uncontrolled arterial hypertension and (2) concentric hypertrophy with left ventricular maximal wall thickness 12mm.

We excluded patients with the following conditions:

Any other causes of LV hypertrophy

Cardiomyopathy

Obstructive coronary heart disease

Severe valvular disease

Inflammatory heart disease

Severe ventricular arrhythmia such as ventricular tachycardia or left bundle branch block

Poor CMR imaging quality

Healthy controls were recruited as volunteers without CVDs (including cardiomyopathy, CAD, severe arrhythmia or conduction block, valvular disease, congenital heart disease and so on) and other organic or systemic diseases on the comprehensive evaluation by patient history, clinical assessment, ECG and echocardiography.

Further information on research design is available in the Nature Portfolio Reporting Summary linked to this article.

More here:
Screening and diagnosis of cardiovascular disease using artificial intelligence-enabled cardiac magnetic resonance ... - Nature.com