ตรวจข้อสอบ > ณัฎฐกานต์ เเสงสิงห์ > Medical & Health Sciences (Secondary Level) | สาขาการแพทย์และสุขภาพ ระดับมัธยมศึกษา > Part 2 > ตรวจ

ใช้เวลาสอบ 2 นาที

Back

# คำถาม คำตอบ ถูก / ผิด สาเหตุ/ขยายความ ทฤษฎีหลักคิด/อ้างอิงในการตอบ คะแนนเต็ม ให้คะแนน
1


How does the concept of “model as a dataset” reshape traditional data-sharing practices in medical imaging?

3. It enables sharing of learned model weights instead of sensitive raw images.

• The concept of “model as a dataset” refers to treating a trained generative AI model itself as a data-sharing resource, rather than sharing the original patient images. • Traditional medical imaging data sharing requires careful handling due to privacy concerns, regulatory restrictions, and patient consent, often making multi-center collaborations complex. • By sharing trained model weights, researchers and institutions can reuse the model to generate synthetic images or perform inference without exposing sensitive raw data. • This approach reduces privacy risks and enables broader collaboration while maintaining regulatory compliance. 1. From Khosravi et al. (2025, The Lancet Digital Health): The article emphasizes that sharing trained generative models can circumvent direct data sharing, preserving patient confidentiality while enabling multi-center research. 2. From Nguyen et al. (2025, Atherosclerotic Cardiovascular Disease Risk Prediction…): While focused on clinical risk models, the article highlights the principle of model-based sharing for regional adaptation—i.e., a model trained on one dataset can be shared and applied elsewhere without exposing the original sensitive clinical data. 7

-.50 -.25 +.25 เต็ม 0 -35% +30% +35%

2


Which analytical conclusion can be drawn about the trade-offs between physics-informed and statistical models?

2. Physics-informed models are more interpretable but computationally intensive.

• Physics-informed models integrate biological or physical principles (e.g., tissue mechanics, MRI physics) into the model. • Pros: The model outputs are more interpretable because they respect known physical laws and anatomical constraints. • Cons: Incorporating these principles makes the model computationally intensive due to complex simulations or constraints. • Statistical models, in contrast, rely on data-driven correlations and often require less domain-specific physics knowledge. They can be faster but may be less interpretable and sometimes violate known physical constraints. This highlights the trade-off: interpretability and realism vs computational efficiency and flexibility. • According to Khosravi et al. (2025), physics-informed generative models are particularly useful when prior knowledge of anatomy or imaging physics is critical, while statistical models excel in data-rich contexts with lower computational costs. • This aligns with the principle of model selection based on domain knowledge and computational constraints in medical image synthesis. 7

-.50 -.25 +.25 เต็ม 0 -35% +30% +35%

3


Why is “mode collapse” considered a critical problem in GAN-based medical image synthesis?

2. It reduces image realism and variety by producing repetitive outputs.

• Mode collapse occurs when the GAN generator produces limited or identical outputs for different inputs, failing to capture the diversity of the training data. • In medical imaging, this is critical because: • Models must generate diverse and realistic synthetic images representing different patients, pathologies, or anatomical variations. • Mode collapse can lead to repetitive images, reducing the utility of synthetic datasets for training AI models, validating algorithms, or educational purposes. • Therefore, addressing mode collapse is essential to ensure high-quality, varied, and clinically useful synthetic images. • According to Khosravi et al. (2025), mode collapse is a major limitation in GANs for medical imaging, as it undermines the generation of diverse pathological or anatomical variations. • This aligns with the principle that generative models in healthcare must preserve diversity to reflect real-world clinical variability, ensuring downstream model performance and reliability. 7

-.50 -.25 +.25 เต็ม 0 -35% +30% +35%

4


Why are healthcare-specific metrics preferred over general-purpose metrics such as FID or SSIM?

2. They better capture clinical accuracy and diagnostic relevance.

• General-purpose metrics like FID (Fréchet Inception Distance) or SSIM (Structural Similarity Index Measure) are designed to quantify visual similarity or realism of images in a generic sense, often using features trained on natural images (e.g., ImageNet). • In medical imaging, visual similarity alone is insufficient. Generated images must also: • Preserve pathologically relevant structures • Reflect accurate clinical features for diagnosis or training AI models • Healthcare-specific metrics are tailored to evaluate clinical fidelity and diagnostic utility, ensuring synthetic images are not only realistic but also clinically meaningful. • According to Khosravi et al., 2025, healthcare-specific evaluation metrics (e.g., lesion detection accuracy, anatomical consistency scores) are crucial because standard computer vision metrics fail to capture subtle pathological or anatomical details. • This aligns with the principle of task-specific evaluation in medical AI, emphasizing that metrics should reflect clinical relevance, not just visual similarity. 7

-.50 -.25 +.25 เต็ม 0 -35% +30% +35%

5


What does the article identify as the key tension between privacy preservation and image fidelity?

1. Higher realism may risk reproducing identifiable patient data.

• Generative models aim to produce high-fidelity synthetic medical images that are realistic enough for clinical use. • However, increasing realism can inadvertently reproduce patterns from the original training data, potentially revealing identifiable patient information. • This creates a key tension: • Privacy preservation: Avoid leaking patient-specific information • Image fidelity: Maintain realistic, diagnostically useful images • Achieving a balance requires careful design of privacy-preserving generative methods, such as differential privacy, federated learning, or sharing model weights instead of raw data. • According to Khosravi et al., 2025, this tension is fundamental in medical image synthesis: maximizing realism may compromise privacy, whereas overly constrained models reduce clinical utility. • Principle: Privacy-preserving AI in healthcare must balance data utility and confidentiality, consistent with ethical and regulatory standards. 7

-.50 -.25 +.25 เต็ม 0 -35% +30% +35%

6


Why is the FDA’s approval of synthetic MRI technology significant for future AI-generated data?

1. It establishes a framework for validating synthetic data equivalence in clinical use.

• FDA clearance of synthetic MRI as image-processing software represents a first regulatory precedent for the clinical use of AI-generated medical images. • This approval: • Confirms that synthetic images can meet quality and reliability standards comparable to real images • Provides a regulatory framework for evaluating the safety, fidelity, and clinical utility of synthetic medical data • Encourages further adoption of AI-generated datasets in research and clinical workflows under regulatory oversight • Without such guidance, adoption of synthetic data in clinical practice would face uncertainty about validation, liability, and patient safety. • According to Khosravi et al., 2025, the FDA’s approval sets an important precedent that allows AI-generated images to be treated as equivalent to real data for certain clinical applications, provided they meet defined validation criteria. • Principle: Regulatory validation of synthetic data is essential for clinical adoption, ensuring both safety and efficacy. 7

-.50 -.25 +.25 เต็ม 0 -35% +30% +35%

7


Which strategy would best mitigate demographic bias in generative models according to the article?

2. Applying diversity-aware training and fairness constraints

• Generative models can inherit biases present in training data, such as overrepresentation of certain demographic groups. • If unaddressed, this leads to biased synthetic images, reducing clinical generalizability and fairness. • Diversity-aware training ensures the model sees balanced samples across demographics, while fairness constraints guide the model to generate images that represent all groups proportionally. • Other strategies (e.g., increasing majority population samples or reducing dataset size) exacerbate bias, while ignoring variation or avoiding external validation risks unreliable and inequitable outputs. • According to Khosravi et al., 2025, mitigating demographic bias is critical for equitable AI in healthcare. • Principle: Fairness in generative models requires both balanced data representation and algorithmic constraints to avoid perpetuating or amplifying existing biases in clinical datasets. 7

-.50 -.25 +.25 เต็ม 0 -35% +30% +35%

8


How do DDPMs exemplify versatility in healthcare image synthesis?

2. They can perform multiple tasks such as denoising, inpainting, and anomaly detection without retraining.

• DDPMs generate images through iterative denoising of noise inputs, learning the underlying data distribution. • This iterative, probabilistic structure allows a single trained DDPM to handle multiple image-based tasks: • Denoising: Remove noise from medical scans • Inpainting: Fill in missing or corrupted regions • Anomaly detection: Highlight deviations from normal anatomy or pathology • Unlike some models requiring retraining for each task, DDPMs can reuse the same model for various applications, showcasing high versatility in healthcare imaging. • According to Khosravi et al., 2025, DDPMs’ flexible reverse diffusion process allows them to adapt to different tasks without retraining, making them valuable for multi-purpose medical image synthesis and analysis. • Principle: Versatility in generative models is enhanced when a model can perform multiple clinically relevant tasks with the same trained weights. 7

-.50 -.25 +.25 เต็ม 0 -35% +30% +35%

9


What analytical insight does the article provide about integrating AI-generated medical images into education and research?

2. It enhances training by providing diverse, realistic datasets without ethical breaches.

• AI-generated synthetic images can augment educational and research datasets by providing: • Diverse cases, including rare pathologies • Realistic images that reflect clinical variability • Ethically safe alternatives to using real patient data, avoiding privacy violations • This allows students, trainees, and researchers to practice and validate AI models without compromising patient confidentiality, improving learning and model robustness. • It does not replace real clinical training entirely, but serves as a complementary tool. • According to Khosravi et al., 2025, synthetic datasets are particularly valuable in medical education and multi-center research where access to large, diverse datasets is limited. • Principle: Synthetic medical images enable scalable, safe, and realistic training and research while respecting privacy and ethical standards. 7

-.50 -.25 +.25 เต็ม 0 -35% +30% +35%

10


Why is regional calibration essential when applying risk prediction models across countries?

2. To adjust for population-specific incidence and lifestyle differences

• Risk prediction models (e.g., Framingham Risk Score) originally developed in Western populations may overestimate or underestimate risk when applied to other populations. • Differences in baseline disease incidence, diet, lifestyle, and genetic factors affect the accuracy of predicted risk. • Regional calibration adjusts model parameters using local epidemiological data, ensuring: • Accurate risk estimation for the target population • Improved clinical decision-making and preventive strategies • Without calibration, models may misclassify patients, leading to over- or under-treatment. • According to Nguyen et al., 2025, East Asian populations have lower baseline ASCVD incidence compared to Western populations, so Western-based models tend to overestimate risk. • Principle: Population-specific calibration is essential for predictive model validity and clinical applicability. 7

-.50 -.25 +.25 เต็ม 0 -35% +30% +35%

11


What analytical conclusion can be drawn when comparing the China-PAR and Framingham models?

2. China-PAR uses local epidemiological data, leading to improved predictive validity.

• The Framingham Risk Score was developed for Western populations and tends to overestimate ASCVD risk in East Asian populations due to differences in baseline incidence, lifestyle, and genetic factors. • The China-PAR model is calibrated using population data from multiple regions in China, which allows it to: • Achieve better predictive validity for East Asian populations • Reduce overestimation of risk • Using population-specific epidemiological data is therefore crucial for the accuracy and reliability of the model. • According to Nguyen et al., 2025, locally developed models such as China-PAR adjust risk predictions appropriately for East Asian populations, whereas Framingham tends to overestimate risk. • Principle: Local calibration enhances predictive accuracy of risk models across populations. 7

-.50 -.25 +.25 เต็ม 0 -35% +30% +35%

12


Based on CVD mortality data, what analytical inference can be made about Japan’s position compared to neighboring countries?

1. Japan’s low CVD mortality suggests effective prevention and healthcare systems.

• Despite differences in population age structures, Japan consistently shows low age-standardized and crude CVD mortality rates. • This indicates that Japan likely benefits from: • Effective preventive strategies (e.g., early detection, public health campaigns) • Advanced healthcare infrastructure and access • Low mortality is not solely due to population size or age distribution, but reflects systematic healthcare effectiveness. • According to Nguyen et al., 2025, Japan’s CVD outcomes outperform neighboring countries like China, Mongolia, and North Korea, highlighting the role of prevention, lifestyle interventions, and healthcare quality. • Principle: Comparative epidemiological analysis can reveal the impact of healthcare systems on population-level disease outcomes. 7

-.50 -.25 +.25 เต็ม 0 -35% +30% +35%

13


What analytical limitation arises when using Western-derived coefficients in East Asian models?

2. It introduces systematic overestimation of ASCVD probability.

• Western-derived models (e.g., Framingham Risk Score) are calibrated using baseline incidence, lifestyle, and risk factor distributions from Western populations. • Applying these coefficients directly to East Asian populations can lead to: • Overestimation of ASCVD risk, because East Asians generally have lower baseline incidence of certain cardiovascular events. • Potential misclassification of patients, resulting in unnecessary interventions or misinformed clinical decisions. • This limitation underscores the importance of population-specific calibration in predictive modeling. • According to Nguyen et al., 2025, using uncalibrated Western coefficients in East Asian cohorts systematically inflates predicted ASCVD probabilities, reducing predictive validity. • Principle: Population-specific calibration ensures accuracy and avoids systematic bias in risk prediction models. 7

-.50 -.25 +.25 เต็ม 0 -35% +30% +35%

14


What policy implication can be derived from country-specific risk models?

1. They allow for targeted national prevention programs.

• Country-specific risk models (e.g., China-PAR, Suita Score, KRPM) are calibrated using local epidemiological data, reflecting: • Population-specific incidence rates • Lifestyle factors such as diet, smoking, and physical activity • With more accurate risk prediction, policymakers can: • Design targeted prevention programs for high-risk groups • Allocate healthcare resources efficiently • Implement population-level interventions tailored to local needs • Using generic models from other regions may misdirect resources or misclassify risk, reducing effectiveness of national health policies. • According to Nguyen et al., 2025, country-specific calibration improves predictive validity, which enables evidence-based national prevention strategies and public health planning. • Principle: Population-specific risk assessment informs precise, effective, and equitable health policy interventions. 7

-.50 -.25 +.25 เต็ม 0 -35% +30% +35%

15


If a model excludes socioeconomic variables, what analytical consequence might occur?

2. Ignored non-biological determinants of disease

• Socioeconomic factors (e.g., income, education, occupation) influence health outcomes and cardiovascular risk beyond biological markers. • Excluding these variables can lead to: • Incomplete risk assessment, as non-biological determinants are ignored • Potential residual confounding, reducing predictive validity for populations with diverse social conditions • Models that include only biological variables may fail to capture the full spectrum of risk, limiting their utility for public health planning and personalized interventions. • According to Nguyen et al., 2025, consideration of socioeconomic determinants is important for accurate ASCVD risk prediction in heterogeneous populations. • Principle: Comprehensive risk models integrate both biological and social determinants to enhance predictive accuracy and equity. 7

-.50 -.25 +.25 เต็ม 0 -35% +30% +35%

16


How might AI improve next-generation ASCVD risk prediction in East Asia?

2. By integrating multimodal data, including imaging and lifestyle information

• Traditional ASCVD risk models primarily rely on demographics, blood tests, and basic clinical variables, which may miss complex patterns. • AI enables integration of multimodal data such as: • Medical imaging (e.g., coronary artery scans) • Lifestyle factors (e.g., diet, physical activity, smoking) • Socioeconomic determinants • This approach allows more personalized and accurate risk predictions, improves population-specific calibration, and supports preventive interventions tailored to East Asian populations. • AI complements rather than replaces traditional models, enhancing predictive power and clinical applicability. • According to Nguyen et al., 2025, next-generation AI models can leverage multimodal datasets to improve ASCVD risk assessment, especially where population characteristics differ from Western cohorts. • Principle: Integrating diverse, population-specific data enhances predictive validity and individualized care. 7

-.50 -.25 +.25 เต็ม 0 -35% +30% +35%

17


What conclusion can be drawn from comparing Mongolia’s and South Korea’s CVD mortality rates?

1. Mortality differences reflect varying effectiveness of national prevention programs.

• CVD mortality rates differ across countries due to differences in healthcare infrastructure, public health initiatives, and preventive strategies. • South Korea has relatively lower CVD mortality, reflecting effective prevention programs, screening, and treatment strategies. • Mongolia, with higher mortality, may have less comprehensive prevention programs and healthcare access limitations. • Thus, comparing mortality rates can highlight the impact of national prevention policies rather than being solely due to demographic or genetic differences. • According to Nguyen et al., 2025, variations in age-standardized CVD mortality among East Asian countries are strongly influenced by the quality and coverage of preventive healthcare programs. • Principle: Epidemiological differences can indicate effectiveness of national health interventions and preventive strategies. 7

-.50 -.25 +.25 เต็ม 0 -35% +30% +35%

18


What is the most logical future direction for improving ASCVD models across East Asia?

1. Establishing multinational data-sharing platforms to harmonize regional models

• Current models in East Asia (e.g., China-PAR, Suita, KRPM) are country-specific, limiting cross-country comparability and generalizability. • Multinational data-sharing platforms would allow: • Integration of large, diverse datasets across East Asian populations • Calibration and harmonization of models for consistent risk prediction • Identification of regional differences in lifestyle, genetics, and healthcare access • This approach enhances model accuracy, predictive validity, and applicability for policy and clinical decision-making. • Relying solely on Western models or ignoring variability would reduce effectiveness and misrepresent risk. • According to Nguyen et al., 2025, collaborative East Asian datasets are critical for developing robust, population-specific, and interoperable ASCVD risk models. • Principle: Cross-national collaboration and data integration improve predictive models while respecting population heterogeneity. 7

-.50 -.25 +.25 เต็ม 0 -35% +30% +35%

19


According to the “image generation trilemma” shown in the figure, what analytical conclusion can be drawn about the relative strengths of VAEs, GANs, and DDPMs in medical image synthesis?

2. GANs provide a balance between image quality and diversity but may suffer from mode collapse.

• The image generation trilemma describes a trade-off among image diversity, quality, and generation speed. • VAEs: tend to produce diverse but lower-quality images, often blurry. • GANs: can generate high-quality, realistic images and maintain moderate diversity, but mode collapse can reduce variety. • DDPMs: produce high-fidelity and diverse images, but are computationally intensive and slower. • Thus, GANs represent a compromise solution, balancing quality and diversity, but are vulnerable to mode collapse, which must be mitigated in medical imaging applications. Based on the figure illustrating the image generation trilemma in the article (ScienceDirect 2025, S258975002500072X): • There are inherent trade-offs among image diversity, quality, and generation speed in generative models. • VAEs tend to prioritize diversity but often produce lower-quality or blurry images. • GANs provide a balance between image quality and diversity, but they can be affected by mode collapse, which reduces variety in generated images. • DDPMs emphasize high-fidelity and realistic image generation, though they require more computation and longer generation times. • Understanding these model-specific strengths and limitations is crucial for selecting the appropriate generative model for medical image synthesis applications. 7

-.50 -.25 +.25 เต็ม 0 -35% +30% +35%

20


Based on Figure, what analytical conclusion can be drawn regarding the distribution of cardiovascular disease (CVD) subtypes across East Asian countries?

1. Ischemic heart disease (IHD) accounts for a higher proportion of CVD deaths in Japan and South Korea compared with China, suggesting regional lifestyle or prevention differences.

• Epidemiological data in the figure indicate that: • Japan and South Korea have a higher proportion of IHD-related deaths. • China has relatively more stroke-related deaths, particularly hemorrhagic stroke. • This suggests that regional differences in diet, lifestyle, hypertension management, and preventive healthcare influence the subtype distribution of CVD. • Understanding these patterns helps tailor public health strategies and prioritize interventions for country-specific CVD burdens. • Based on ScienceDirect 2025, S2772374725000511, Figure comparing CVD subtypes in East Asian countries: • Highlights variation in IHD vs. stroke proportions among Japan, South Korea, and China. • Principle: Regional lifestyle, diet, and healthcare policies shape CVD subtype prevalence. 7

-.50 -.25 +.25 เต็ม 0 -35% +30% +35%

ผลคะแนน 125.3 เต็ม 140

แท๊ก หลักคิด
แท๊ก อธิบาย
แท๊ก ภาษา