โจทย์ ปรนัย
What is the primary goal of the article according to its introduction?

The primary goal of the article is to synthesise or link Generative AI in Medical Imaging such as oversampling the minoritized group in order to close the fairness gap.

In the ASCVD article, it is stated that American ASCVD risk calculators like PCE and FRS significantly overestimate risk in East Asian people, causing China, Japan, and the Republic of Korea to make their own risk calculators since they have a higher risk of having stroke but low risk of having CHD. In the Generative AI article, it is stated that DDPMs (Diffusion Denoising Probabilistic Models) which have both high quality and diversity although the speed is slow, it could be applied in X-rays as it makes the image clearer. This shows that both articles are linked together in order to increase the generalizability or disaggregated data for East Asian Americans.

โจทย์ ปรนัย
How do generative AI models differ from traditional discriminative models in healthcare applications?

Generative AI models can create their content and new synthetic data which solves specific medical or science tasks as its core function is synthesis while traditional discriminative models in healthcare applications give predictions and make decisions since its core function is interpretation and decision-making.

Generative AI differs from traditional discriminative models in healthcare applications because it focuses on data creation rather than interpretation. For example, traditional discriminative factors look at factors like sex or cholesterol in order to calculate ASCVD patient risk in the next ten years. Generative AI focuses on data creation, enabling the creation of synthetic datasets which helps offers solutions to limitations such as solving bias (fairness gap) by oversampling the minoritized group.

โจทย์ ปรนัย
What is meant by the term “model as a dataset”?

Generative AI have a concept called as "a model as a dataset", making it learn patterns of original data and store it in Weights (internal parameters). Sharing weights enables generation of new synthetic images for use in anatomical structures or pathological correlations.

Based on the Generative AI article, sharing these trained weights allows other generative AI to generate new synthetic images that resembles the original data. This makes it able to generate data with patient privacy solutions for sharing medical data, as the generated images may look or contain similar properties to the original data but without replicating the patient's sensitive data.

โจทย์ ปรนัย
Which statement correctly distinguishes physics-informed and statistical models?

By using two models (Physics-informed and statistical models), synthetic datasets are generated. Physics-Informed models offer high interpretability though it require expertise and resources while statistical models learn directly from data patterns.

Physics-Informed models are rule-based approaches that incorporate specific knowledge or physical principle such as math equations, stimulating biological principles. On the other hand, statistical models learn from patterns and distributions, categorized into three main models: VAEs, GANs, and DDPMs. These three models especially DDPMs can be applied in image transformation ( uses VAE and GAN) and stress-testing models (uses DDPM).

โจทย์ ปรนัย
According to the article, what does the “image generation trilemma” describe?

Image generation trilemma involves high sample quality, comprehensive mode coverage (diversity), and rapid sampling rates (speed) although most shifts focus on diversity and quality.

Based on both articles, statistical models (VAE, GAN, DDPM) can be used in imaging and applied in ASCVD like using advanced DDPMS to forecast disease progression such as brain tumor. Looking at statistical models, VAEs consist of encoder (add all 500 page essay into one page known as the latent space) and decoder (reconstruct and summarize) which work together in order to provide high speed but low quality or diversity. GANs consists of the competition between the generator and discriminator as the situation acts as a robber finding ways to steal stuff (generator) and a police learning about the way robbers plan or think, making it easier for the police to evaluate robbers (discriminator). This makes its quality high but low diversity due to mode collapse. DDPMs add noise to clear image by markov chain and learns to reverse the process by denoising it using the gradient descent. Its quality and diversity is high but the speed is rather slow. This shows that by combining all three models form the image generation trilemma (quality, diversity, and speed).

โจทย์ ปรนัย
What is the Human Turing Test used for in medical image synthesis?

The Human Turing test is a process of checking the image quality of generated images.

Experts with high knowledge with different experience levels are asked to discern between real and generated medical images in order to check the standard and quality of AI medical images.

โจทย์ ปรนัย
Which of the following is NOT mentioned as a potential benefit of synthetic data in healthcare?

The benefits of synthetic data in healthcare are privacy preservation, versatility across tasks, enhancing data diversity and size, eliminating potential biases, indentification of source dataset using guidelines, and collaboration.

It is said that challenges like data copying, biases, identification of source dataset, and interpretability can be solved by using different privacy techniques, diversity-aware sampling, standardized reporting guidelines, and developing relevant interpretability metrics.

โจทย์ ปรนัย
What is one major ethical concern associated with generative AI in medical imaging?

Data copying and patient reidentification is one of the major concerns associated with generative AI in medical imaging as generative models risk revealing sensitive patient information if the images are close to the original data.

In order to reduce the tendency of creating similar images to the original data, generative AI can anonymizes sensitive patient information by generating realistic images contains similar properties of biological characteristics of real patient data.

โจทย์ ปรนัย
What regulatory precedent did the article cite for synthetic data technologies?

The article specifically talks about the FDA's clearance of synthetic MRI

MRI technologies were classified as image processing software rather than as completely novel modalities as the FDA required external validation which demonstrate that the radiologist's diagnosis remained the same when using both synthetic images and conventional images

โจทย์ ปรนัย
What is the main purpose of the article?

It is clearly stated that American / Western models significantly overestimate risk in East Asian populations and generative AI models such as oversampling and risk prediction models which could be applied in East Asia.

FRS and PCE (western models) significantly overestimate East Asian risk as East Asian people have higher rates of stroke but lower rates of Coronary Heart Disease (CHD) compared to Western people. This causes China, Japan, and Korea to make their own risk calculators but there are two main problems: generalizability within population or country and accuracy if it could be applied and used for East Asian immigrants living in Western countries.

โจทย์ ปรนัย
Which of the following models was originally developed for a Western population?

All of these models except FRS (Framingham risk score) are risk calculators and models made by East Asian countries (China, Japan, and Korea). Framingham risk score of The Framingham CHD risk greatly overestimated CHD risk for East Asian cohorts as East Asian persons have lower rates of CHD but higher rates of stroke.

Western models like FRS and PCE significantly overestimate the risk for East Asian persons due to lack of representation in founding cohorts, founded or developed in US community, few East Asian subjects, and distinct disease patterns ( westerns tend to have high rates of CHD and low rates of stroke but East Asians have lower rates of CHD and higher rates of stroke)

โจทย์ ปรนัย
Why might Western-based risk prediction models overestimate ASCVD risk in East Asian populations?

Inaccuracy of risk in East Asian persons by Western models (PCE or FRS) are mostly due to different disease patterns.

East Asians have lower CHD/ASCVD rates but higher stroke mortality rates. Comparing the founding cohorts shows that the 10-year CHD event rates were 8.0% in Framingham men and 2.8% in women, but only 1.5% and 0.6% in the Chinese Multi-provincial Cohort Study (CMCS) men and women.

โจทย์ ปรนัย
What is the key advantage of the China-PAR model compared to Western-based models?

China - PAR model is made in China and the data are collected from Chinese population so the results are less likely to be overestimated and it also consists of both genetic and lifestyle factors.

Lifestyle factors are such as smoking and WC, which is related to lifestyle and central obesity. China-PAR cohort creates the prediction tool (the lifetime ASCVD risk), aiming to motivate lifestyle modifications. Genetic factors (family history of ASCVD)are published for the China-PAR risk equations include family history of ASCVD which is the strongest genetic risk.

โจทย์ ปรนัย
Which of the following variables is not typically included in ASCVD risk prediction models discussed in the article?

Age, blood pressure, cholesterol, and Smoking status are all included in all risk calculators.

Although some models (like China-PAR and Japanese models) include family history of ASCVD, it is inadequate to show that core risk equations use genetic ancestry markers to calculate the 10-year risk score.

โจทย์ ปรนัย
What is a major difference between the Suita Score and the Framingham Risk Score?

Suita Score was made from Japanese cohort studies and developed in response to the limitation of Western models, such as the Framingham Risk Score (FRS) and the ACC/AHA's Pooled Cohort Equation (PCE), which significantly overestimate ASCVD risk in East Asian populations.

Suita score was selected from different published risk prediction scores in Japan because of its ability to accurately estimate the absolute incidence of Coronary Heart Disease (CHD) and incorporates risk factors such as age, smoking, sex, high-density lipoprotein cholesterol (HDL-C), low-density lipoprotein cholesterol (LDL-C), and family history of CHD.

โจทย์ ปรนัย
According to the article, what is a potential benefit of developing East Asia–specific risk models?

โจทย์ ปรนัย
Which factor was highlighted as influencing ASCVD risk differences among East Asian countries?

โจทย์ ปรนัย
What future direction does the article suggest for improving ASCVD risk prediction?

โจทย์ ปรนัย
Which statement best explains the key difference in how VAEs, GANs, and DDPMs generate medical images according to the figure?

DDPMs add noise to clear image and then do the reverse process (denoising) while VAEs and GANs use encoder-decoder or generator-discriminator .

VAEs consist of encoder (add all 500 page essay into one page known as the latent space) and decoder (reconstruct and summarize) which work together in order to provide high speed but low quality or diversity. GANs consists of the competition between the generator and discriminator as the situation acts as a robber finding ways to steal stuff (generator) and a police learning about the way robbers plan or think, making it easier for the police to evaluate robbers (discriminator). This makes its quality high but low diversity due to mode collapse. DDPMs add noise to clear image by markov chain and learns to reverse the process by denoising it using the gradient descent. Its quality and diversity is high but the speed is rather slow.

โจทย์ ปรนัย
Which of the following best explains the trend shown in Figure comparing age-standardized and crude CVD mortality rates among East Asian countries?