| 1 |
A pneumothorax classifier performs well on the hospital’s original test set. Researchers then use a generative model to remove chest tubes from pneumothorax radiographs while preserving the visible pneumothorax. The classifier’s accuracy falls sharply.
Which interpretation is most strongly supported by this result?
|
The classifier had learned to use the chest tube as a shortcut associated with pneumothorax rather than relying sufficiently on the pathological features. |
|
The important clue is that the pneumothorax itself was preserved, but the chest tube was removed and accuracy dropped dramatically. That suggests the classifier had been relying on the chest tube as a shortcut rather than sufficiently recognizing pneumothorax itself. |
The article apecifically reports that a pneumothorax classifier's accuracy fropped from 93.3% to 17.9% when chest tubes were artificially removed. It decribed this work as using counterfactuas to evaluate possible shortcuts and their effect on model performance. |
7 |
-.50
-.25
+.25
เต็ม
0
-35%
+30%
+35%
|
| 2 |
A tumour-segmentation algorithm requires T1, T2, post-contrast T1, and FLAIR MRI sequences. A generative model is proposed to synthesize a missing FLAIR sequence when it was not acquired.
Which study design would provide the strongest evidence that the synthetic sequence is clinically useful?
|
Compare tumour-segmentation performance using synthetic FLAIR images with performance using complete real MRI sequences on an independent patient set. |
|
The question asked whether the synthetic sequence is clinically useful not merely whther it looks realistic. The strongest evidence is whther it helps the actual clinical task, which is tumor segmentation, on patients who were not used for development. |
The article report that synthetic missing MRi sequences improved tumor segmentation, it also exphasized testing synthetic images in practical clinical tasks such as disease classification and combining computational evaluation with clinical validation. |
7 |
-.50
-.25
+.25
เต็ม
0
-35%
+30%
+35%
|
| 3 |
A treatment-aware diffusion model predicts future brain tumour appearance. Its predictions are most reliable within approximately four months, while similarity to actual outcomes decreases substantially at longer intervals.
How should clinicians use this model most responsibly?
|
Use the model mainly within the validated short-term period and communicate increasing uncertainty for longer-term forecasts. |
|
The question tells us that the model is reliable for about four months and becomes less familiar to actual outcomes later. This means that clinicians should not treat long term predictions as a certain. The four month reliability figure itself comes from the question, not the article. But B is the answer most consistent with the article's emphasis on uncertainty and responsible clinical use. |
The evidence says that advanced DDPMs can generate potential future brain tumor progression and describes their potential use for prognosis and clinical decision making. The article's future directions also emphasizes uncertainty quantiification for stochastic prediction models. |
7 |
-.50
-.25
+.25
เต็ม
0
-35%
+30%
+35%
|
| 4 |
An unsupervised diffusion model is trained on a large collection of unlabelled pelvic radiographs. Researchers later use its internal features to train a segmentation system with only 20 expert-annotated images.
What does this application most clearly demonstrate?
|
Generative models can provide reusable image representations for downstream tasks beyond image synthesis. |
|
The model wasn't simply generating images. Researchers used its internal features to improve another task which is segmentation with only 20 annotated images. |
The article specifically reports that using internal features of a generative model enhanced pelvis radiagraph segmentaton by using only annotated samples. It also explains that generative models can have value beyond image generation and can support multiple downstream applications. |
7 |
-.50
-.25
+.25
เต็ม
0
-35%
+30%
+35%
|
| 5 |
A generative model produces synthetic brain MRIs that closely resemble its training images. Investigators discover that some training scans were duplicated several times in the source dataset.
Which response best addresses the privacy concern identified in the article?
|
Curate and deduplicate the source data, assess similarity and membership-inference risk, and consider differential privacy or post-hoc anonymisation. |
|
The duplicated training scans increase concern about the model memorizing or reproducing patient information. I don't think deleting names is enough. |
The article says data copying can occur when multiple copies of an image exist in the training dataset. It recommends careful data curation and assessment of anonymization and reidentification risks. |
7 |
-.50
-.25
+.25
เต็ม
0
-35%
+30%
+35%
|
| 6 |
A research group releases a synthetic chest-radiograph dataset generated from a pretrained model. The group does not disclose the model’s original training sources, prompt conditions, inference settings, or post-processing procedures.
What is the most important scientific problem with this release?
|
Researchers cannot adequately assess potential bias, reproduce the dataset, or determine how generation decisions influenced the outputs. |
|
If researchers don't know the original training data, prompts, settings, or processing steps, they can't properly understand where the synthetic images came from or reproduce the results. |
The article says that identifying the datasets used to train generative models can be challenging and that it hingers assesment of potential biases or limitations. It recommends standardized reporting guidelines, dataset fingerprinting, and trusted validation. The article also mentioned that understanding how generative models learn and generate data is necessary for trust and interpretability. |
7 |
-.50
-.25
+.25
เต็ม
0
-35%
+30%
+35%
|
| 7 |
Two models generate synthetic lung images. Model A achieves a better Fréchet inception distance calculated with an ImageNet-pretrained network. However, radiologists find that Model A frequently distorts small pulmonary lesions that are important for diagnosis.
Which conclusion is most appropriate?
|
A general-purpose feature extractor may emphasise global visual similarity while failing to capture clinically important local abnormalities. |
|
FID is a computational image quality metric but the article warns that standard metrics don't necessary represent clinical relevance or diagnostic utility. |
The article explicitly says standard metrics can fail to reflect clinical relevance or diagnostic utility and recommends combining computational approaches with clinical validation. The exact claim about an ImageNet-pretrained network emphasizing global rather than local features isn't directly stated in the article. But the overall conclusion that a general image metric should not be treated as proof of clinical superiority is strongly supported. |
7 |
-.50
-.25
+.25
เต็ม
0
-35%
+30%
+35%
|
| 8 |
Researchers want to simulate patient-specific blood flow from vascular images. The outputs must follow known fluid-dynamic principles while also adapting to patterns found in real patient data.
Which future modelling strategy best matches this objective?
|
Combine physics-informed constraints with a statistical generative model trained on patient data. |
|
The task requires both known physical laws and patterns learned from real patients. |
The article explicitly identifies hybrid models combining physics informed and statistical approaches. They incorporate domain specific knowledge and constraints as an important future direction. |
7 |
-.50
-.25
+.25
เต็ม
0
-35%
+30%
+35%
|
| 9 |
A company develops software that creates synthetic high-quality MRI images from faster, lower-quality acquisitions. It seeks approval for routine clinical use.
Which evidence package is most consistent with the regulatory pathway discussed in the article?
|
Proof that radiologists achieve equivalent diagnostic performance with synthetic and conventional images, supported by standardised tasks, multiple readers, and postmarket monitoring |
|
For routine clinical approval, merely showing that images look good isn't enough. The key question is whether clinicians can actually make equivalent diagnoses using the synthetic images |
The article says the FDA has cleared synthetic MRI technologies as image processing software and required extensive clinical validation showing equivalent radiologist diagnostic performance compared with conventional images. It then describes the pathway as standardized diagnostic tasks and rigorous clinical validation. |
7 |
-.50
-.25
+.25
เต็ม
0
-35%
+30%
+35%
|
| 10 |
A 42-year-old Chinese man has a low predicted 10-year ASCVD risk because of his age. However, he has persistently elevated blood pressure, smokes, and has a high predicted lifetime risk.
Which interpretation best reflects the purpose of lifetime-risk models described in the article?
|
Lifetime-risk estimation can identify younger individuals who may benefit from earlier lifestyle modification despite low short-term risk. |
|
A 42-year-old may have a low 10-year risk simply because they are young, even though risk factors can create substantial risk over their lifetime. |
The ASCVD article says Chinese guidelines incorporated lifetime risk estimation into risk classification. |
7 |
-.50
-.25
+.25
เต็ม
0
-35%
+30%
+35%
|
| 11 |
An East Asian American patient has a body mass index of 28 kg/m². A health survey defines obesity using a threshold of 30 kg/m² developed primarily for non-Asian populations.
What is the most likely consequence?
|
The survey may underestimate obesity-related cardiovascular and metabolic risk in East Asian populations. |
|
An East Asian person with BMI 28 would already meet the article's Asian specific obesity threshold, even though they don't meet the conventional 30 kg/m^2 threshold. |
The article says Asian people can develop central obesity and related diabetes, metabolic syndrome, and CVD at lower BMIs than non-Asian people. It states that obesity cut offs have been lowered to less than or equal to 27.5 or 25.0 kg/m^2 rather than 30 kg/m^2. |
7 |
-.50
-.25
+.25
เต็ม
0
-35%
+30%
+35%
|
| 12 |
A national study reports one average prevalence of diabetes and hypertension for all “Asian Americans,” combining East Asian, South Asian, and Southeast Asian participants.
Why is this result insufficient for developing an ASCVD model for Korean Americans?
|
Aggregation may conceal substantial differences in risk-factor prevalence among Asian subgroups. |
|
“Asian American” is not one homogeneous risk group. East Asian, South Asian, and Southeast Asian populations have different rates of hypertension, diabetes, obesity, and dyslipidemia. |
The article gives different risk factor rates for East Asian, South Asian, and Southeast Asian immigrants. For example, diabetes was 6.3% in East Asian immigrants in contrast to 8.6% in South Asian and 11.3% in Southeast Asian immigrants. |
7 |
-.50
-.25
+.25
เต็ม
0
-35%
+30%
+35%
|
| 13 |
The AHA PREVENT calculator removes race and ethnicity as explicit predictors because their effects may be partly reflected in socioeconomic variables. The model has not been specifically validated in East Asian populations.
Which conclusion is most scientifically appropriate?
|
The calculator’s validity in East Asian populations remains uncertain until its discrimination and calibration are directly evaluated. |
|
Removing race does not automatically prove that a calculator works equally well in every population. |
The article says prevent removed race or ethnicity because their effects may already be reflected in socioeconomic data, but its accuracy has not been specifically validated in East Asian people. |
7 |
-.50
-.25
+.25
เต็ม
0
-35%
+30%
+35%
|
| 14 |
A Korean patient has several major modifiable cardiovascular risk factors, but a prediction model assigns a low absolute 10-year risk because ASCVD incidence is relatively low in the source population.Why do Korean guidelines recommend using a risk score mainly as an additional risk enhancer in low- or moderate-risk groups rather than relying on it alone?
|
Low population event rates may produce a low calculated risk even when clinically important risk factors require treatment. |
|
A prediction score depends partly on the underlying event rate of the population. If the baseline incidence is relatively low, someone's calculated absolute risk can remain low despite having important risk factors. |
The article says Korean incidence of CHD is low and that Western derived calculators have overestimated absolute risk in Korea. It also notes that the Korean guidelines use the Korean risk score as a risk enhancer in low or moderate risk groups. |
7 |
-.50
-.25
+.25
เต็ม
0
-35%
+30%
+35%
|
| 15 |
When SCORE2 and SCORE2-OP are applied in Korea, risk is underestimated in adults aged 40–59 years but overestimated in older adults.
Which revision would most directly address this problem?
|
Perform age-specific recalibration and validate the revised estimates in contemporary Korean cohorts. |
|
The article says SCORE2/SCORE2-OP behaves differently across age groups in Korea and underestimating risk in ages 40–59 and overestimating risk in older individuals. |
The article explicitly reports this age dependent miscalibration and says it highlights the need for a region specific calculator. It also emphasizes the importance of current local data and validation. |
7 |
-.50
-.25
+.25
เต็ม
0
-35%
+30%
+35%
|
| 16 |
A deep-learning system estimates coronary artery calcium from retinal photographs. Its predictions are comparable with CT-measured CAC for cardiovascular-event prediction in one Korean study.
Which conclusion is best supported?
|
The method is promising as a potentially cost-effective and radiation-free approach, but broader validation is required before routine adoption. |
|
The Korean study found retinal deep learning estimates of CAC comparable to CT measured CAC for predicting CVD events. But the article specifically says this should only be considered an alternative if further validated. |
The article says the retinal approach is comparable to CT-measured CAC and could become a cost effective, radiation free alternative if further validated. |
7 |
-.50
-.25
+.25
เต็ม
0
-35%
+30%
+35%
|
| 17 |
Researchers evaluate the predictive value of coronary CT angiography using patients who were referred for imaging because clinicians already suspected coronary disease. They then propose applying the findings to all asymptomatic adults.
What is the principal methodological concern?
|
The study population may have a higher baseline risk than the general asymptomatic population, creating selection bias. |
|
If researchers only study people who were referred for coronary imaging because clinicians already suspected disease, that group is likely to have a different or higher risk than the general asymptomatic population. Applying those results to everyone could therefore be very misleading. |
This is warned about in the ASCVD article which is imaging based risk modifiers such as CAC, carotid ultrasound, and ABI may be affected by selection bias because many patients receiving these imaging studies may already be at higher ASCVD risk. |
7 |
-.50
-.25
+.25
เต็ม
0
-35%
+30%
+35%
|
| 18 |
Several biomarkers and physiological measurements improve risk classification in individual Japanese studies. However, they are not routinely collected in clinics and have not been adequately validated in population-based cohorts.
Which policy is most evidence-based?
|
Require evidence of incremental predictive value, population-level validation, feasibility, and cost-effectiveness before routine incorporation. |
|
The answer is C because the article explains that several biomarkers showed improved risk classification in individual Japanese studies, but that does not automatically mean they are ready to be added to a national risk calculator. |
The article points out that some biomarkers can improve prediction when added to traditional risk factors, but they have not been adequately studied in population based cohorts and some are not routinely measured in clinical practice. |
7 |
-.50
-.25
+.25
เต็ม
0
-35%
+30%
+35%
|
| 19 |
A hospital is considering two separate applications:
• Application A: Generating a large, offline dataset of rare disease images for training diagnostic classifiers, where both image quality and representation of diverse disease patterns are essential.
• Application B: Producing rapid, low-stakes anatomical previews during an educational simulation, where generation speed is more important than achieving the highest possible image quality.
Based on the figure and the article, which deployment strategy is most appropriate?
|
Use DDPMs for Application A and VAEs for Application B, while separately validating whether the outputs are medically accurate for each intended use. |
|
Application A requires high image quality and diverse disease patterns for training diagnostic classifiers, which aligns directly with DDPMs located at the quality diversity vertex of the image generation trilemma shown in the diagram. Conversely, Application B prioritizes generation speed over maximum image quality for rapid, low stakes educational previews, making VAEs the ideal choice due to their positioning near the speed axis. |
The image generation trilemma maps out the triangle between quality, diversity, and speed for different architectures, showing GANs along the quality speed edge, VAEs near the speed diversity base, and DDPMs along the quality diversity edge. Based on these architectural trade offs, DDPMs successfully avoid the mode collapse and limited diversity of GANs to provide the dataset needed for rare disease classification, while VAEs enable the rapid sampling necessary for interactive previews. |
7 |
-.50
-.25
+.25
เต็ม
0
-35%
+30%
+35%
|
| 20 |
An East Asian patient has elevated body mass index, chronic kidney disease, low physical activity, and a family history of coronary artery disease. Blood pressure, cholesterol, diabetes status, and tobacco use are entered identically into both the AHA PREVENT and ACC/AHA PCE assessment systems.
The two systems classify the patient differently. Which explanation is most consistent with the figure?
|
AHA PREVENT directly includes body mass index and chronic kidney disease in its calculation, whereas ACC/AHA PCE excludes body mass index and treats chronic kidney disease, family history, and physical activity as additional risk enhancers rather than core predictors. |
|
I chose B because the figure shows that the two calculators handle these risk factors differently. In the AHA PREVENT column, BMI and chronic kidney disease are included in the risk calculator, while in the ACC/AHA PCE column, these factors are not included as core calculator variables but instead, the figure identifies several of them as risk enhancers or modifeirs. This explains why the same patient can receive different risk classifications even when blood pressure, cholesterol, diabetes, and tobacco use are entered identically. |
The article also supports this idea by explaining that different cardiovascular risk models have differences in their predictors and that risk enhancers can affect risk assessment. It emphasizes that many risk models also have limitations in validation and relevancy, meaning their results cannot automatically be assumed to be identical across populations. |
7 |
-.50
-.25
+.25
เต็ม
0
-35%
+30%
+35%
|