Q.1 What is the primary imaging modality used for visualizing bone structures in orthopedic diagnostics?
MRI
CT
Ultrasound
X-ray
Explanation - X-ray imaging produces high-contrast images of dense bone structures and is the most common modality for bone assessment.
Correct answer is: X-ray
Q.2 Which of the following is a key difference between supervised and unsupervised learning?
Supervised learning uses labeled data; unsupervised does not
Unsupervised learning requires a neural network; supervised does not
Supervised learning can only classify; unsupervised can only regress
Unsupervised learning uses a loss function; supervised does not
Explanation - Supervised algorithms learn from labeled examples, whereas unsupervised algorithms discover patterns in unlabeled data.
Correct answer is: Supervised learning uses labeled data; unsupervised does not
Q.3 In medical image segmentation, what does the Dice coefficient measure?
Accuracy of edge detection
Overlap between predicted and ground truth masks
Sensitivity of the model
Training speed of the algorithm
Explanation - The Dice coefficient quantifies spatial overlap; a value of 1 indicates perfect match.
Correct answer is: Overlap between predicted and ground truth masks
Q.4 Which machine learning model is commonly used for predicting patient readmission risk from electronic health records?
Support Vector Machine
Random Forest
Convolutional Neural Network
K-means Clustering
Explanation - Random Forests handle tabular EHR data well and provide feature importance for interpretability.
Correct answer is: Random Forest
Q.5 What preprocessing step is essential before feeding CT images into a deep learning model?
Color normalization
Intensity scaling to Hounsfield units
Histogram equalization
Edge sharpening
Explanation - CT images are measured in Hounsfield units; standardizing this scale improves model consistency.
Correct answer is: Intensity scaling to Hounsfield units
Q.6 Which loss function is most suitable for imbalanced segmentation tasks?
Mean Squared Error
Cross‑Entropy
Focal Loss
Huber Loss
Explanation - Focal loss down‑weights easy examples and focuses training on hard, often minority class pixels.
Correct answer is: Focal Loss
Q.7 What is the main advantage of using a 3D CNN over a 2D CNN for volumetric medical imaging?
Faster inference time
Ability to capture inter‑slice context
Reduced number of parameters
Easier training with limited data
Explanation - 3D CNNs process all spatial dimensions simultaneously, capturing volume‑wide features.
Correct answer is: Ability to capture inter‑slice context
Q.8 Which metric is best for evaluating a binary classification model's ability to avoid false positives?
Precision
Recall
Accuracy
F1 Score
Explanation - Precision measures the proportion of true positives among all positives predicted, thus reflecting false positive rate.
Correct answer is: Precision
Q.9 What does the abbreviation 'GAN' stand for in deep learning?
Gradient Augmented Network
Generative Adversarial Network
Gaussian Activation Network
Graph Attention Network
Explanation - GANs consist of a generator and discriminator trained in opposition to produce realistic synthetic data.
Correct answer is: Generative Adversarial Network
Q.10 Which imaging modality uses magnetic fields and radio waves to generate images of soft tissues?
X-ray
CT
MRI
Ultrasound
Explanation - MRI relies on nuclear magnetic resonance to produce detailed soft‑tissue contrast.
Correct answer is: MRI
Q.11 In a convolutional neural network, what is the purpose of the activation function?
To increase the depth of the network
To introduce non‑linearity
To reduce the number of channels
To perform dimensionality reduction
Explanation - Activation functions enable networks to model complex, non‑linear relationships.
Correct answer is: To introduce non‑linearity
Q.12 Which of the following best describes 'transfer learning'?
Training a model from scratch on a large dataset
Using a pre‑trained model as a starting point for a new task
Transferring labels between datasets
Moving data from one imaging modality to another
Explanation - Transfer learning fine‑tunes a model already trained on a large dataset to adapt it to a new, often smaller, task.
Correct answer is: Using a pre‑trained model as a starting point for a new task
Q.13 Which deep learning architecture is commonly used for image segmentation tasks?
AlexNet
ResNet
U‑Net
VGG16
Explanation - U‑Net’s encoder‑decoder structure with skip connections is tailored for pixel‑wise segmentation.
Correct answer is: U‑Net
Q.14 In medical image classification, what does the term 'augmented reality' refer to?
Overlaying diagnostic information onto real‑time video
Synthetic image generation for training
Using virtual reality to train surgeons
Augmenting the resolution of low‑quality images
Explanation - Augmented reality enhances the real world with computer‑generated content for better situational awareness.
Correct answer is: Overlaying diagnostic information onto real‑time video
Q.15 What is the primary function of a radiomics pipeline?
To segment tumors automatically
To extract quantitative imaging features for analysis
To reconstruct 3D models from 2D images
To encrypt medical images for security
Explanation - Radiomics converts images into high‑dimensional data that can be correlated with outcomes.
Correct answer is: To extract quantitative imaging features for analysis
Q.16 Which of the following is a common artifact seen in ultrasound imaging?
Beam hardening
Motion blur
Speckle noise
Photon starvation
Explanation - Speckle is a granular interference pattern inherent to coherent imaging like ultrasound.
Correct answer is: Speckle noise
Q.17 What does 'AUC' stand for in the context of ROC analysis?
Area Under Curve
Average Utility Coefficient
Accuracy Under Confidence
Augmented Utility Curve
Explanation - AUC measures the overall ability of a classifier to discriminate between classes across thresholds.
Correct answer is: Area Under Curve
Q.18 Which of the following is NOT a typical preprocessing step for MRI data before machine learning?
Bias field correction
Skull stripping
Histogram equalization
Normalization to zero mean, unit variance
Explanation - Histogram equalization is rarely used for MRI; intensity standardization and bias correction are preferred.
Correct answer is: Histogram equalization
Q.19 What is the purpose of data augmentation in training deep learning models for medical imaging?
To reduce the training time
To artificially increase dataset size and improve generalization
To convert 2D images to 3D
To lower the model complexity
Explanation - Augmentation applies random transformations to create more diverse training examples.
Correct answer is: To artificially increase dataset size and improve generalization
Q.20 Which of the following best describes a 'saliency map' in medical image interpretation?
A heatmap highlighting regions that influence a model’s prediction
A map of image gradients over time
A spatial map of pixel intensities
A visualization of the loss function
Explanation - Saliency maps reveal which areas contribute most to the output, aiding interpretability.
Correct answer is: A heatmap highlighting regions that influence a model’s prediction
Q.21 In the context of deep learning, what does 'dropout' aim to prevent?
Data leakage
Overfitting
Underfitting
Vanishing gradients
Explanation - Dropout randomly deactivates neurons during training, forcing the network to learn robust features.
Correct answer is: Overfitting
Q.22 Which imaging modality is best suited for functional brain imaging?
CT
MRI (fMRI)
X-ray
Ultrasound
Explanation - Functional MRI measures blood oxygen level changes, revealing brain activity patterns.
Correct answer is: MRI (fMRI)
Q.23 What does 'ROI' stand for in medical imaging analysis?
Region of Interest
Rate of Imaging
Radiographic Output Index
Randomized Observation Interval
Explanation - ROI defines a specific area in an image that is targeted for analysis or annotation.
Correct answer is: Region of Interest
Q.24 Which technique is commonly used to register a pre‑operative CT scan to a real‑time ultrasound during surgery?
Affine transformation
Rigid registration
Non‑rigid (deformable) registration
Intensity normalization
Explanation - Non‑rigid registration accounts for tissue deformations that occur intra‑operatively.
Correct answer is: Non‑rigid (deformable) registration
Q.25 What is the main limitation of using CNNs for small medical datasets?
High computational cost
Difficulty in feature extraction
Risk of overfitting
Inability to handle color images
Explanation - CNNs have many parameters; without enough data they tend to memorize training samples.
Correct answer is: Risk of overfitting
Q.26 Which of the following best describes 'batch normalization'?
Normalizing input image sizes
Normalizing outputs of each layer during training
Normalizing batch labels
Normalizing the loss function
Explanation - Batch normalization stabilizes and accelerates training by keeping activations within a similar range.
Correct answer is: Normalizing outputs of each layer during training
Q.27 Which evaluation metric is most appropriate for a segmentation task with highly imbalanced classes?
Pixel Accuracy
Jaccard Index (IoU)
Mean Absolute Error
Top‑1 Accuracy
Explanation - IoU focuses on overlap and is less influenced by the majority class.
Correct answer is: Jaccard Index (IoU)
Q.28 Which of the following is an example of a 'clinical trial' involving AI for medical imaging?
A study comparing AI‑based detection of diabetic retinopathy to human graders
A survey of radiologists’ preferences for imaging devices
A retrospective review of imaging protocols
A benchmark dataset publication
Explanation - Clinical trials evaluate AI algorithms in a prospective, regulated setting.
Correct answer is: A study comparing AI‑based detection of diabetic retinopathy to human graders
Q.29 Which of the following is a common challenge when applying deep learning to chest X‑ray datasets?
Large number of modalities
High resolution images
Class imbalance and label noise
Limited GPU memory
Explanation - Rare pathologies lead to imbalanced classes; noisy labels arise from expert disagreements.
Correct answer is: Class imbalance and label noise
Q.30 What does 'semantic segmentation' aim to achieve?
Detecting object boundaries
Classifying each pixel into a category
Extracting image features
Generating captions for images
Explanation - Semantic segmentation assigns a class label to every pixel in an image.
Correct answer is: Classifying each pixel into a category
Q.31 In a convolutional layer, what does the term 'kernel size' refer to?
Number of output channels
Size of the convolutional filter
Stride length
Number of layers in the network
Explanation - Kernel size indicates the spatial extent (e.g., 3×3) of the filter applied to the input.
Correct answer is: Size of the convolutional filter
Q.32 Which of the following best describes 'attention mechanism' in transformer models?
Weighting input pixels based on relevance
Skipping layers during training
Adjusting learning rate dynamically
Normalizing gradients
Explanation - Attention assigns importance scores to input elements to focus processing on crucial parts.
Correct answer is: Weighting input pixels based on relevance
Q.33 What is the primary advantage of using a 'pre‑trained' model for medical image analysis?
Increased interpretability
Reduced training time
Guaranteed higher accuracy
Elimination of preprocessing steps
Explanation - Pre‑trained models provide useful feature representations, reducing the amount of data needed to train.
Correct answer is: Reduced training time
Q.34 Which type of neural network is most suitable for time‑series analysis of ECG signals?
Convolutional Neural Network
Recurrent Neural Network
Feedforward Neural Network
Generative Adversarial Network
Explanation - RNNs are designed to capture temporal dependencies in sequential data like ECG.
Correct answer is: Recurrent Neural Network
Q.35 In a supervised classification task, which metric would you use to evaluate performance on highly imbalanced data?
Accuracy
Precision‑Recall AUC
Mean Squared Error
Cross‑Entropy Loss
Explanation - PR‑AUC focuses on the minority class performance and is more informative than ROC‑AUC in imbalance.
Correct answer is: Precision‑Recall AUC
Q.36 What does the term 'data curation' refer to in medical imaging datasets?
Collecting images from scanners
Cleaning and labeling data, removing duplicates
Training models on curated data
Storing data in encrypted format
Explanation - Curation ensures that the dataset is accurate, consistent, and ready for machine learning.
Correct answer is: Cleaning and labeling data, removing duplicates
Q.37 Which of the following best describes 'k‑fold cross‑validation'?
Splitting data into k subsets and training on each subset
Training on the first k samples and testing on the rest
Using k different models to vote
Normalizing the dataset into k parts
Explanation - K‑fold CV trains on k‑1 folds and validates on the remaining fold, rotating through all folds.
Correct answer is: Splitting data into k subsets and training on each subset
Q.38 In the context of image registration, what does the term 'rigid transformation' refer to?
Translation only
Translation and rotation only
Translation, rotation, and scaling
Non‑linear deformation
Explanation - A rigid transformation preserves distances and angles, allowing only translation and rotation.
Correct answer is: Translation and rotation only
Q.39 Which of the following is an example of a 'semi‑supervised' learning approach?
Using only labeled data for training
Using labeled data and unlabeled data together
Training with a single labeled example
Training without any labels
Explanation - Semi‑supervised learning leverages both labeled and unlabeled samples to improve performance.
Correct answer is: Using labeled data and unlabeled data together
Q.40 What is the main benefit of using '3‑D Slicer' in medical imaging research?
It provides a deep learning library
It offers 3‑D visualization and segmentation tools
It encrypts patient data
It automatically diagnoses diseases
Explanation - 3‑D Slicer is an open‑source platform for image analysis, visualization, and segmentation.
Correct answer is: It offers 3‑D visualization and segmentation tools
Q.41 Which loss function is commonly used for image classification tasks?
Dice Loss
Triplet Loss
Cross‑Entropy Loss
Huber Loss
Explanation - Cross‑entropy measures the difference between predicted probabilities and true labels.
Correct answer is: Cross‑Entropy Loss
Q.42 What does 'overfitting' refer to in machine learning?
The model generalizes well to new data
The model performs poorly on training data
The model memorizes training data and fails to generalize
The model underestimates the variance
Explanation - Overfitting occurs when a model captures noise in the training set, reducing performance on unseen data.
Correct answer is: The model memorizes training data and fails to generalize
Q.43 In the context of medical imaging, what is a 'ground truth' annotation?
Automatically generated labels by an algorithm
Labels derived from the training loss
Expert‑verified labels considered correct
Randomly assigned labels
Explanation - Ground truth refers to the most accurate labels provided by clinicians or specialists.
Correct answer is: Expert‑verified labels considered correct
Q.44 Which imaging modality is most commonly used for detecting bone fractures in emergency settings?
MRI
CT
X‑ray
Ultrasound
Explanation - X‑ray is fast, widely available, and highly effective for visualizing bone fractures.
Correct answer is: X‑ray
Q.45 Which type of neural network is most effective for modeling sequences like respiratory waveforms?
Convolutional Neural Network
Recurrent Neural Network
Graph Neural Network
Autoencoder
Explanation - RNNs capture dependencies over time, making them suitable for waveform analysis.
Correct answer is: Recurrent Neural Network
Q.46 Which of the following is NOT a typical hyperparameter in a CNN?
Learning rate
Batch size
Dropout probability
Color depth
Explanation - Color depth is an image property, not a model hyperparameter.
Correct answer is: Color depth
Q.47 What does 'data imbalance' refer to?
Unequal distribution of training and validation sets
Unequal number of samples per class
Unequal image resolutions
Unequal number of layers in a model
Explanation - Data imbalance occurs when some classes have many more examples than others.
Correct answer is: Unequal number of samples per class
Q.48 Which of these is a typical evaluation metric for detection models?
Intersection over Union (IoU)
Mean Absolute Error
F1 Score
Precision
Explanation - IoU measures how well predicted bounding boxes overlap with ground truth.
Correct answer is: Intersection over Union (IoU)
Q.49 Which of the following best describes 'k‑means clustering'?
A supervised classification algorithm
An unsupervised algorithm grouping data into k clusters
A deep learning model for segmentation
A technique for dimensionality reduction
Explanation - k‑means partitions data into k clusters by minimizing intra‑cluster variance.
Correct answer is: An unsupervised algorithm grouping data into k clusters
Q.50 Why is 'bias field correction' applied to MR images?
To enhance edge detection
To correct intensity non‑uniformity due to inhomogeneous magnetic fields
To remove motion artifacts
To reduce noise using Gaussian smoothing
Explanation - Bias field causes spatially varying intensity that must be removed for accurate analysis.
Correct answer is: To correct intensity non‑uniformity due to inhomogeneous magnetic fields
Q.51 What is the main goal of 'model calibration' in medical AI?
To adjust model predictions to match observed probabilities
To reduce the number of parameters
To increase model speed
To augment data
Explanation - Calibration ensures that predicted risk scores reflect true outcome probabilities.
Correct answer is: To adjust model predictions to match observed probabilities
Q.52 Which of the following is a common source of noise in PET imaging?
Thermal noise
Poisson noise due to photon counting statistics
Speckle noise
Quantization noise
Explanation - PET signals arise from discrete photon events, leading to Poisson noise.
Correct answer is: Poisson noise due to photon counting statistics
Q.53 Which term describes a network that learns to generate new data resembling a training set?
Classifier
Autoencoder
Generator
Detector
Explanation - Generators are part of GANs and produce synthetic data.
Correct answer is: Generator
Q.54 What is the function of a 'decoder' in an encoder‑decoder architecture?
Compress the input data
Reconstruct the input or produce an output from latent representation
Select relevant features
Classify the input data
Explanation - The decoder reconstructs high‑dimensional outputs from compact latent features.
Correct answer is: Reconstruct the input or produce an output from latent representation
Q.55 Which evaluation metric is most sensitive to class imbalance in binary classification?
Accuracy
Recall
Specificity
Precision
Explanation - Recall (sensitivity) focuses on the proportion of positives correctly identified, often highlighted in imbalanced scenarios.
Correct answer is: Recall
Q.56 In medical imaging, what is 'intensity normalization' primarily used for?
To reduce computational complexity
To ensure that pixel values are comparable across scans
To increase image resolution
To convert color images to grayscale
Explanation - Normalization brings diverse image intensities onto a common scale, improving model performance.
Correct answer is: To ensure that pixel values are comparable across scans
Q.57 What is the main advantage of using a 'U‑Net' architecture over a plain CNN for segmentation?
Fewer parameters
Skip connections that preserve spatial detail
Ability to process 3D volumes
Reduced training time
Explanation - Skip connections combine encoder features with decoder layers, enhancing localization accuracy.
Correct answer is: Skip connections that preserve spatial detail
Q.58 Which of the following is a common data augmentation technique for retinal fundus images?
Random cropping and rotation
Spectral filtering
Wavelet decomposition
Histogram equalization
Explanation - Geometric augmentations such as cropping and rotating simulate variations in patient positioning.
Correct answer is: Random cropping and rotation
Q.59 What does the 'learning rate' hyperparameter control?
The number of neurons in each layer
The speed at which weights are updated during training
The depth of the network
The regularization strength
Explanation - A higher learning rate moves weights faster; too high can cause divergence.
Correct answer is: The speed at which weights are updated during training
Q.60 Which of the following best describes 'model interpretability'?
The ability of a model to run on any hardware
The ease with which humans can understand why a model made a decision
The speed of inference
The accuracy of predictions
Explanation - Interpretability allows clinicians to trust and validate AI outputs.
Correct answer is: The ease with which humans can understand why a model made a decision
Q.61 In CT imaging, what is the primary source of radiation exposure to the patient?
X‑ray photons
Magnetic fields
Ultrasound waves
Infrared light
Explanation - CT uses a rotating X‑ray tube that emits photons to produce cross‑sectional images.
Correct answer is: X‑ray photons
Q.62 Which of the following is a typical application of transfer learning in medical imaging?
Training a model from scratch on a small dataset
Fine‑tuning a model pre‑trained on ImageNet for tumor segmentation
Using only unsupervised learning
Applying unsupervised clustering to raw images
Explanation - Transfer learning leverages learned features from a large dataset to improve performance on a related task.
Correct answer is: Fine‑tuning a model pre‑trained on ImageNet for tumor segmentation
Q.63 What is the primary benefit of '3D convolution' over '2D convolution' in volumetric data?
Less memory usage
Capturing spatial context across slices
Simplified architecture
Faster training
Explanation - 3D convolutions process volume data simultaneously, preserving inter‑slice relationships.
Correct answer is: Capturing spatial context across slices
Q.64 Which type of regularization adds a penalty term proportional to the square of the magnitude of weights?
L1 Regularization
Dropout
L2 Regularization
Early Stopping
Explanation - L2 (Ridge) regularization discourages large weights by adding a squared‑norm penalty.
Correct answer is: L2 Regularization
Q.65 What is the main purpose of using a 'confusion matrix' in classification?
To compute the loss function
To visualize model predictions vs actual labels
To reduce the model size
To augment the dataset
Explanation - The confusion matrix shows true positives, false positives, true negatives, and false negatives.
Correct answer is: To visualize model predictions vs actual labels
Q.66 Which of the following is an advantage of using 'deep learning' over traditional hand‑crafted feature extraction in medical imaging?
Requires less computational resources
Automatically learns relevant features from data
Guarantees interpretability
Always provides higher accuracy
Explanation - Deep models discover hierarchical features directly from raw images.
Correct answer is: Automatically learns relevant features from data
Q.67 In medical imaging, what does the term 'noise floor' refer to?
The lowest possible intensity value
The baseline level of noise inherent to the imaging system
The threshold for detecting lesions
The maximum intensity of a pixel
Explanation - The noise floor indicates the background noise level, affecting image quality.
Correct answer is: The baseline level of noise inherent to the imaging system
Q.68 Which of these is a common technique for addressing class imbalance during training?
Using larger batch sizes
Class weighting
Increasing dropout rate
Applying data augmentation only to majority class
Explanation - Assigning higher weights to minority classes penalizes misclassifications more heavily.
Correct answer is: Class weighting
Q.69 What is the 'gradient descent' algorithm used for?
Classifying images
Optimizing model parameters by minimizing a loss function
Augmenting data
Normalizing image intensities
Explanation - Gradient descent iteratively updates weights in the direction of the negative gradient.
Correct answer is: Optimizing model parameters by minimizing a loss function
Q.70 Which of the following best describes '3D U‑Net'?
An extension of U‑Net for volumetric data
A 2D CNN with three input channels
A 3‑layer MLP for segmentation
A 3D GAN architecture
Explanation - 3D U‑Net applies 3‑D convolutions to process entire volumes.
Correct answer is: An extension of U‑Net for volumetric data
Q.71 In deep learning, what is the purpose of a 'batch normalization' layer?
To normalize input image sizes
To stabilize learning by normalizing layer inputs
To increase the number of parameters
To enforce sparsity
Explanation - Batch normalization reduces internal covariate shift and speeds up convergence.
Correct answer is: To stabilize learning by normalizing layer inputs
Q.72 Which of the following is a common imaging modality for evaluating cardiovascular function?
CT
MRI
X‑ray
Ultrasound
Explanation - Cardiac MRI provides high‑contrast images of myocardial tissue and motion without ionizing radiation.
Correct answer is: MRI
Q.73 What does the 'soft‑max' function do in a neural network?
Converts logits to probabilities that sum to 1
Applies a threshold to produce binary outputs
Reduces overfitting by dropout
Normalizes input pixel values
Explanation - Soft‑max turns raw scores into a probability distribution over classes.
Correct answer is: Converts logits to probabilities that sum to 1
Q.74 Which of the following best describes 'active learning'?
A method where the model selects the most informative samples for labeling
A training technique that uses only labeled data
A form of unsupervised clustering
A type of transfer learning
Explanation - Active learning prioritizes labeling of examples that will improve model performance most.
Correct answer is: A method where the model selects the most informative samples for labeling
Q.75 What is the main advantage of using a 'residual block' in a deep CNN?
Simplifies the architecture
Enables training of deeper networks by mitigating vanishing gradients
Reduces the number of parameters
Increases model interpretability
Explanation - Residual connections provide shortcut paths that help gradients flow backward.
Correct answer is: Enables training of deeper networks by mitigating vanishing gradients
Q.76 Which type of imaging is best suited for evaluating soft‑tissue contrast in the brain?
CT
MRI
X‑ray
Ultrasound
Explanation - MRI provides superior soft‑tissue contrast compared to CT or X‑ray.
Correct answer is: MRI
Q.77 Which loss function is particularly useful when dealing with class imbalance in segmentation?
Mean Squared Error
Dice Loss
Cross‑Entropy Loss
Huber Loss
Explanation - Dice loss directly optimizes the Dice coefficient, favoring minority class segmentation.
Correct answer is: Dice Loss
Q.78 Which of the following is an example of a 'supervised' deep learning task in medical imaging?
Clustering MRI scans
Detecting tumors in labeled X‑ray images
Generating synthetic ultrasound images
Learning a latent space representation of CT scans
Explanation - Supervised tasks use labeled examples for training.
Correct answer is: Detecting tumors in labeled X‑ray images
Q.79 What is the purpose of 'image registration' in multimodal imaging?
To adjust image contrast
To align images from different modalities into a common coordinate system
To reduce noise
To segment tissues
Explanation - Registration ensures that anatomical structures overlap across modalities.
Correct answer is: To align images from different modalities into a common coordinate system
Q.80 In a medical imaging pipeline, what role does 'data augmentation' play?
It reduces the computational cost
It generates new data to improve generalization
It normalizes intensities
It removes artifacts
Explanation - Augmentation creates varied examples, helping models generalize beyond the training set.
Correct answer is: It generates new data to improve generalization
Q.81 Which of the following is NOT a typical characteristic of medical imaging data?
Large dimensionality
High class imbalance
Low resolution
High annotation cost
Explanation - Medical images are often high resolution to capture fine anatomical details.
Correct answer is: Low resolution
Q.82 What does the term 'clinical validation' refer to in AI for medical imaging?
Testing the algorithm on synthetic data
Evaluating algorithm performance in a real clinical setting
Performing unit tests on code
Running the model on a benchmark dataset
Explanation - Clinical validation assesses the model's effectiveness in real-world practice.
Correct answer is: Evaluating algorithm performance in a real clinical setting
Q.83 Which of the following is a major concern when deploying AI models in healthcare?
Lack of GPU support
High interpretability
Privacy and data security
Unlimited data availability
Explanation - Patient data must be protected to comply with regulations like HIPAA.
Correct answer is: Privacy and data security
Q.84 In a 2‑D CNN, what effect does a larger stride have on the output feature map?
Increases spatial resolution
Decreases spatial resolution
Increases number of channels
Has no effect
Explanation - A larger stride skips more input positions, reducing the size of the output map.
Correct answer is: Decreases spatial resolution
Q.85 What is the main advantage of using a 'pre‑trained' CNN for medical image classification?
It eliminates the need for any fine‑tuning
It provides a starting point that can generalize better with limited data
It guarantees higher accuracy than a model trained from scratch
It reduces inference time by 90%
Explanation - Pre‑trained weights capture low‑level features that transfer across tasks.
Correct answer is: It provides a starting point that can generalize better with limited data
Q.86 Which of these is a typical output of a 'segmentation' model?
Bounding boxes around objects
Probability map of disease presence
Pixel‑wise class labels
Feature vectors
Explanation - Segmentation assigns a class to each pixel, producing a mask.
Correct answer is: Pixel‑wise class labels
Q.87 Which term refers to the practice of using synthetic data to augment training datasets?
Data synthesis
Data curation
Data annotation
Data compression
Explanation - Synthetic data generation creates new realistic samples for training.
Correct answer is: Data synthesis
Q.88 Which imaging modality typically requires the least patient preparation?
MRI
CT
X‑ray
Ultrasound
Explanation - Ultrasound usually only requires a gel and no fasting or contrast agents.
Correct answer is: Ultrasound
Q.89 Which of the following best describes 'binary cross‑entropy loss'?
Used for multi‑class classification
Used for binary classification to measure difference between predicted probability and actual label
Used for regression tasks
Used for image reconstruction
Explanation - Binary cross‑entropy quantifies error for two‑class predictions.
Correct answer is: Used for binary classification to measure difference between predicted probability and actual label
Q.90 What is the primary purpose of using a 'Gaussian filter' in image preprocessing?
Sharpen edges
Reduce noise by smoothing
Enhance contrast
Increase resolution
Explanation - Gaussian filtering averages neighboring pixels, reducing high‑frequency noise.
Correct answer is: Reduce noise by smoothing
Q.91 Which of the following is a key component of a 'deep learning inference pipeline'?
Model training on GPU
Data augmentation during inference
Model loading and GPU allocation
Labeling new data
Explanation - Inference involves loading a trained model and allocating compute resources.
Correct answer is: Model loading and GPU allocation
Q.92 In medical imaging, which of the following is a commonly used dataset for training lung nodule detection algorithms?
ImageNet
LIDC‑IDRI
MNIST
CIFAR‑10
Explanation - LIDC‑IDRI provides CT scans with annotated lung nodules for research.
Correct answer is: LIDC‑IDRI
Q.93 What does the 'Adam' optimizer adjust during training?
The learning rate manually
The gradient descent step using momentum and adaptive learning rates
The number of epochs
The model architecture
Explanation - Adam combines RMSProp and momentum for efficient optimization.
Correct answer is: The gradient descent step using momentum and adaptive learning rates
Q.94 Which of the following is a typical challenge when performing automated breast cancer detection in mammography?
High image contrast
Large amount of labeled data
Class imbalance with few malignant cases
Uniform tissue density
Explanation - Malignant lesions are rare, leading to imbalanced datasets.
Correct answer is: Class imbalance with few malignant cases
Q.95 Which of the following best describes 'data drift' in AI models?
Changes in patient demographics
Gradual changes in input data distribution over time
Immediate change in model weights
Loss of training data
Explanation - Data drift refers to the shift in distribution that can degrade model performance.
Correct answer is: Gradual changes in input data distribution over time
Q.96 What is the primary function of a 'residual connection' in a neural network?
To add regularization
To provide a shortcut for gradient flow
To reduce overfitting
To increase feature dimensionality
Explanation - Residual connections help mitigate vanishing gradients in deep networks.
Correct answer is: To provide a shortcut for gradient flow
Q.97 Which of the following is an example of a medical imaging application that uses 'graph neural networks'?
Brain network connectivity analysis
Tumor segmentation in MRI
Edge detection in X‑ray
Noise removal in ultrasound
Explanation - GNNs are used to model connectivity patterns between brain regions.
Correct answer is: Brain network connectivity analysis
Q.98 What does 'batch size' control during training?
Number of epochs
Number of training examples processed before updating weights
Number of layers in the network
Learning rate
Explanation - Batch size determines how many samples are used in one gradient update.
Correct answer is: Number of training examples processed before updating weights
Q.99 Which of the following is a characteristic of 'semi‑supervised learning' in medical imaging?
Only labeled data is used
Only unlabeled data is used
Both labeled and unlabeled data are used to improve performance
Data is never labeled
Explanation - Semi‑supervised methods exploit unlabeled data alongside labeled examples.
Correct answer is: Both labeled and unlabeled data are used to improve performance
Q.100 Which of the following best describes 'image augmentation' techniques?
Adding random noise to images
Applying transformations such as rotation, scaling, flipping, and cropping
Normalizing image intensities
Converting images to grayscale
Explanation - Augmentation increases dataset diversity by manipulating images.
Correct answer is: Applying transformations such as rotation, scaling, flipping, and cropping
Q.101 Which of the following metrics is best for evaluating a model that predicts the probability of disease presence?
Accuracy
AUC‑ROC
F1 Score
Mean Squared Error
Explanation - AUC‑ROC measures model discrimination across thresholds, suitable for probabilistic outputs.
Correct answer is: AUC‑ROC
Q.102 What is the primary benefit of using a 'pre‑trained' model for a new medical imaging task?
Zero training time
Immediate perfect accuracy
Faster convergence and better generalization with limited data
No need for validation data
Explanation - Transfer learning leverages learned representations, accelerating training.
Correct answer is: Faster convergence and better generalization with limited data
Q.103 Which of the following is a key factor influencing the choice of imaging modality for liver disease?
Patient age
Imaging cost
Tissue contrast and radiation dose
Color depth of image
Explanation - MRI offers superior soft‑tissue contrast with no ionizing radiation, important for liver imaging.
Correct answer is: Tissue contrast and radiation dose
Q.104 Which of the following is a common preprocessing step for chest X‑ray images?
Resizing to a fixed resolution
Adding synthetic noise
Color to grayscale conversion
Both A and C
Explanation - Resizing ensures uniform input size; X‑ray images are already grayscale.
Correct answer is: Both A and C
Q.105 What is the primary function of a 'decoder' in a U‑Net architecture?
Encode images to a latent vector
Upsample feature maps and merge with encoder features
Classify images into disease categories
Extract features for classification
Explanation - The decoder reconstructs segmentation masks using skip connections.
Correct answer is: Upsample feature maps and merge with encoder features
Q.106 Which of the following is NOT a typical step in training a CNN for medical image segmentation?
Data augmentation
Patch extraction
Label smoothing
Overfitting to training data
Explanation - Overfitting is a problem to avoid, not a training step.
Correct answer is: Overfitting to training data
Q.107 What does the 'soft‑max' function output?
Binary predictions
Probabilities for each class that sum to 1
Logits
Feature vectors
Explanation - Soft‑max normalizes logits to a probability distribution over classes.
Correct answer is: Probabilities for each class that sum to 1
Q.108 Which of the following best describes 'model ensembling'?
Using a single model with more layers
Combining predictions from multiple models to improve performance
Training a model on a larger dataset
Reducing model complexity
Explanation - Ensembling averages or votes across models, often boosting accuracy.
Correct answer is: Combining predictions from multiple models to improve performance
Q.109 Which of the following is a key advantage of using 3‑D imaging over 2‑D slices?
Higher resolution
Simplified preprocessing
Better capture of spatial context and continuity
Reduced storage requirements
Explanation - 3‑D data preserves volumetric relationships between slices.
Correct answer is: Better capture of spatial context and continuity
Q.110 What does 'cross‑validation' help prevent when training machine learning models?
Overfitting
Underfitting
Data leakage
Both A and C
Explanation - Cross‑validation provides a robust estimate of generalization and helps detect leakage.
Correct answer is: Both A and C
Q.111 Which of the following is a common application of deep learning in retinal imaging?
Diabetic retinopathy screening
Bone fracture detection
Brain tumor segmentation
Cardiac rhythm analysis
Explanation - Deep models automatically detect micro‑aneurysms and hemorrhages in retinal photographs.
Correct answer is: Diabetic retinopathy screening
Q.112 Which of the following best describes 'gradient clipping'?
Limiting the maximum value of gradients to prevent exploding gradients
Normalizing gradients to zero mean
Increasing the learning rate after each epoch
Reducing the number of gradient updates per batch
Explanation - Gradient clipping caps gradients to avoid instability during training.
Correct answer is: Limiting the maximum value of gradients to prevent exploding gradients
Q.113 What is the purpose of a 'heatmap' in the context of medical imaging AI?
To display pixel intensities
To visualize attention or saliency of the model
To encode segmentation masks
To represent the loss landscape
Explanation - Heatmaps highlight image regions most influential in the model's decision.
Correct answer is: To visualize attention or saliency of the model
Q.114 Which of the following best defines 'radiomics'?
Radiological imaging in the era of AI
Extraction of quantitative features from medical images
Radiation therapy planning
Use of radiologists for AI supervision
Explanation - Radiomics transforms images into high‑dimensional feature vectors for predictive modeling.
Correct answer is: Extraction of quantitative features from medical images
Q.115 Which metric is most suitable for evaluating segmentation of small lesions?
Pixel Accuracy
Dice Coefficient
Mean Absolute Error
Precision
Explanation - Dice emphasizes overlap and is less biased by majority classes, making it ideal for small lesions.
Correct answer is: Dice Coefficient
Q.116 What is the main role of an 'encoder' in a U‑Net?
Upsample feature maps
Extract hierarchical features by reducing spatial resolution
Predict class probabilities
Generate synthetic images
Explanation - The encoder compresses spatial information into deeper representations.
Correct answer is: Extract hierarchical features by reducing spatial resolution
Q.117 Which of the following is NOT typically used to evaluate a classification model?
Confusion matrix
Loss curve
Radiation dose
Precision
Explanation - Radiation dose is a clinical metric, not a model evaluation metric.
Correct answer is: Radiation dose
Q.118 Which of these imaging modalities is most commonly used for functional brain mapping?
CT
MRI (fMRI)
X‑ray
Ultrasound
Explanation - Functional MRI measures blood oxygen changes, indicating brain activity.
Correct answer is: MRI (fMRI)
Q.119 Which technique can be used to prevent overfitting in deep learning models?
Dropout
Increasing batch size
Removing activation functions
Training for more epochs
Explanation - Dropout randomly disables neurons during training, reducing over‑reliance on specific weights.
Correct answer is: Dropout
Q.120 What does 'hyper‑parameter tuning' involve?
Optimizing model architecture automatically
Manually selecting the best hyper‑parameters through validation
Adjusting the labels in the dataset
Cleaning the data
Explanation - Hyper‑parameter tuning searches for optimal values like learning rate, batch size, etc.
Correct answer is: Manually selecting the best hyper‑parameters through validation
Q.121 What is the primary reason for using a '3D U‑Net' for brain tumor segmentation?
To process 2‑D slices individually
To capture volumetric context across slices
To reduce model size
To perform edge detection
Explanation - Brain tumors are 3‑D structures; a 3‑D U‑Net leverages spatial continuity.
Correct answer is: To capture volumetric context across slices
Q.122 Which of the following is a characteristic of 'unsupervised learning'?
Uses labeled training data
Finds patterns without labels
Requires a large labeled dataset
Always requires a supervised loss function
Explanation - Unsupervised methods discover structure directly from data without ground truth.
Correct answer is: Finds patterns without labels
Q.123 Which of the following best describes 'foveal‑centered' imaging?
Imaging the peripheral retina
Focusing imaging on the central visual field for high resolution
Imaging only the optic nerve
Using infrared light for retinal imaging
Explanation - Foveal‑centered imaging captures the central retina where visual acuity is highest.
Correct answer is: Focusing imaging on the central visual field for high resolution
Q.124 In the context of medical AI, what does 'bias' refer to?
Systematic error favoring certain groups or outcomes
Random noise in image acquisition
The weight initialization method
The learning rate schedule
Explanation - Bias can arise from training data or model design, leading to unequal performance.
Correct answer is: Systematic error favoring certain groups or outcomes
Q.125 What is a major ethical consideration for AI in medical imaging?
Ensuring the model runs fast on GPUs
Guaranteeing 100% accuracy
Protecting patient privacy and data security
Increasing model complexity
Explanation - AI systems must comply with regulations like HIPAA to safeguard personal health information.
Correct answer is: Protecting patient privacy and data security
Q.126 Which of the following is a common use of 'GANs' in medical imaging?
Segmentation of organs
Generating synthetic medical images for data augmentation
Classifying tumors
Detecting motion artifacts
Explanation - GANs can produce realistic images to enlarge limited datasets.
Correct answer is: Generating synthetic medical images for data augmentation
Q.127 In a 2‑D CNN, what happens when the filter size is increased?
The receptive field grows, capturing more context
The model becomes deeper automatically
The number of parameters decreases
The output resolution increases
Explanation - Larger filters aggregate information over a bigger spatial area.
Correct answer is: The receptive field grows, capturing more context
Q.128 Which of the following is a typical application of 'object detection' in medical imaging?
Locating tumors in CT scans
Segmenting the entire organ
Generating synthetic images
Predicting patient age
Explanation - Object detection outputs bounding boxes around lesions or abnormalities.
Correct answer is: Locating tumors in CT scans
Q.129 What does 'image registration' aim to achieve?
Align images from different modalities into a common space
Standardize image intensities
Convert images to grayscale
Reduce image size
Explanation - Registration ensures that corresponding anatomical points coincide.
Correct answer is: Align images from different modalities into a common space
Q.130 What is the role of a 'soft‑max layer' at the output of a classification model?
To produce a probability distribution over classes
To perform dimensionality reduction
To apply dropout
To normalize input features
Explanation - Soft‑max converts raw scores to probabilities that sum to 1.
Correct answer is: To produce a probability distribution over classes
Q.131 Which of the following is a standard preprocessing step for PET images?
Gaussian smoothing
Bias field correction
Co‑registration to CT
Both A and C
Explanation - Smoothing reduces noise; co‑registration aligns PET with anatomical CT.
Correct answer is: Both A and C
Q.132 Which metric is often used to evaluate a model’s calibration?
Confusion matrix
Reliability diagram
Mean squared error
Dice coefficient
Explanation - Reliability diagrams plot predicted probabilities against observed frequencies.
Correct answer is: Reliability diagram
Q.133 What does the term 'over‑parameterization' refer to in deep learning?
Using too few parameters for a problem
Using more parameters than necessary, risking overfitting
Using parameters that are too large in magnitude
Using a single parameter for all weights
Explanation - Over‑parameterized models have a high capacity relative to the training data.
Correct answer is: Using more parameters than necessary, risking overfitting
Q.134 Which of the following is a primary reason to use a 'learning rate scheduler'?
To keep the learning rate constant
To reduce the learning rate gradually to improve convergence
To increase the model size
To generate synthetic data
Explanation - Schedulers help refine training by decreasing the step size over time.
Correct answer is: To reduce the learning rate gradually to improve convergence
Q.135 Which type of medical imaging is best suited for high‑resolution imaging of lung tissue?
X‑ray
CT
MRI
Ultrasound
Explanation - CT provides high‑resolution cross‑sectional images of lung parenchyma.
Correct answer is: CT
Q.136 In medical imaging, why is 'class imbalance' a frequent issue?
Rare diseases result in few positive samples
All classes are equally represented
There is always abundant data
Imaging devices produce balanced data
Explanation - Diseases are often under‑represented compared to normal anatomy.
Correct answer is: Rare diseases result in few positive samples
Q.137 Which of the following best defines 'normalization' in the context of deep learning?
Reducing the size of the dataset
Scaling input values to a standard range
Increasing the number of layers
Adding random noise
Explanation - Normalization improves training stability and convergence.
Correct answer is: Scaling input values to a standard range
Q.138 Which of the following is a common artifact seen in MRI?
Motion artifacts
Speckle noise
Beam hardening
Photon starvation
Explanation - Patient movement during scan causes ghosting or blurring in MRI images.
Correct answer is: Motion artifacts
Q.139 Which of the following best describes 'semi‑automatic segmentation'?
Fully automated segmentation
Manual segmentation with no assistance
Segmentation aided by user interaction or constraints
Segmentation using only image intensity
Explanation - The user provides guidance such as seeds or bounding boxes.
Correct answer is: Segmentation aided by user interaction or constraints
Q.140 Which of the following is NOT a typical evaluation metric for classification models?
Accuracy
Sensitivity
Precision
MSE (Mean Squared Error)
Explanation - MSE is used for regression, not classification.
Correct answer is: MSE (Mean Squared Error)
Q.141 Which of the following best describes 'data bias' in medical AI?
Unequal representation of patient subgroups in the training data
The model outputs biased predictions for all patients
The dataset is too large
All training images are perfectly labeled
Explanation - Bias arises when certain populations are under‑represented, skewing model performance.
Correct answer is: Unequal representation of patient subgroups in the training data
Q.142 What is the primary objective of a 'loss function' in training neural networks?
To determine the architecture of the model
To measure the difference between predictions and ground truth
To decide the batch size
To visualize the training data
Explanation - The loss function quantifies training errors, guiding weight updates.
Correct answer is: To measure the difference between predictions and ground truth
Q.143 Which of the following best describes 'data augmentation'?
Increasing the dataset size by adding synthetic or transformed samples
Removing outliers from the dataset
Splitting data into training and test sets
Normalizing image intensities
Explanation - Augmentation applies random transformations to expand the data pool.
Correct answer is: Increasing the dataset size by adding synthetic or transformed samples
Q.144 What is a typical preprocessing step for MRI data before feeding into a CNN?
Skull stripping
Color normalization
Adding Gaussian noise
Downscaling to 64×64
Explanation - Removing non‑brain tissue focuses the model on relevant anatomy.
Correct answer is: Skull stripping
Q.145 Which of the following is an example of a 'clinical decision support system'?
A CNN that classifies skin lesions
An AI model that recommends treatment plans based on imaging
A database of medical images
A software for image compression
Explanation - Decision support systems provide actionable recommendations to clinicians.
Correct answer is: An AI model that recommends treatment plans based on imaging
Q.146 What does 'train‑validation‑test split' aim to prevent?
Data leakage
Model complexity
GPU over‑utilization
Loss function over‑fitting
Explanation - Separating data prevents the model from seeing test samples during training.
Correct answer is: Data leakage
Q.147 Which of the following best describes a 'radiologist‑in‑the‑loop' workflow?
The radiologist automatically labels all images
The AI system makes final diagnoses without human input
The radiologist reviews AI outputs and makes the final decision
The radiologist uses AI only for image storage
Explanation - This collaborative approach combines AI efficiency with human expertise.
Correct answer is: The radiologist reviews AI outputs and makes the final decision
Q.148 Which of the following is a characteristic of a 'deep neural network'?
Only one hidden layer
Multiple layers that transform data hierarchically
No activation functions
Deterministic weights
Explanation - Deep networks consist of many stacked layers extracting increasingly abstract features.
Correct answer is: Multiple layers that transform data hierarchically
Q.149 Which of the following is a major challenge in applying deep learning to retinal imaging?
High resolution required
Low contrast of the retinal layers
Limited availability of labeled data
All of the above
Explanation - Retinal imaging demands high precision, has subtle contrast, and labeled data is scarce.
Correct answer is: All of the above
Q.150 What does 'beam hardening' affect in CT images?
Spatial resolution
Contrast between soft tissues
Quantitative accuracy of Hounsfield units
Image noise
Explanation - Beam hardening causes artifacts that bias intensity values.
Correct answer is: Quantitative accuracy of Hounsfield units
Q.151 Which of the following is a common post‑processing step after segmentation?
Applying morphological operations to clean up the mask
Running the model again on the same image
Changing the input resolution
Normalizing the output intensities
Explanation - Morphological filtering removes small spurious segments and smooths boundaries.
Correct answer is: Applying morphological operations to clean up the mask
Q.152 Which of the following is NOT an advantage of deep learning over traditional computer vision techniques?
Automatic feature extraction
Better scalability to large datasets
Require fewer computational resources
Better performance on complex tasks
Explanation - Deep learning typically demands more GPU memory and compute than classical methods.
Correct answer is: Require fewer computational resources
Q.153 Which of the following best describes 'early stopping'?
Stopping training when validation loss starts to increase
Stopping training after a fixed number of epochs
Stopping training when training loss is zero
Stopping training when GPU memory is full
Explanation - Early stopping prevents overfitting by halting training at the optimal point.
Correct answer is: Stopping training when validation loss starts to increase
Q.154 Which of the following is a standard evaluation metric for a detection model?
Mean Average Precision (mAP)
Dice Coefficient
Precision
Accuracy
Explanation - mAP measures detection accuracy across classes and thresholds.
Correct answer is: Mean Average Precision (mAP)
Q.155 What does 'adversarial training' aim to achieve?
Improve robustness against small input perturbations
Reduce model size
Increase training speed
Generate synthetic images
Explanation - Adversarial training exposes the model to perturbed examples to make it more resilient.
Correct answer is: Improve robustness against small input perturbations
Q.156 Which of the following is an example of a 'clinical endpoint' in medical AI research?
Area under the ROC curve
Time to correct diagnosis
Model accuracy
Training loss
Explanation - Clinical endpoints measure real‑world patient outcomes or workflow improvements.
Correct answer is: Time to correct diagnosis
Q.157 Which of the following best defines 'model interpretability'?
The model's ability to run on low‑end hardware
The extent to which humans can understand why the model made a decision
The speed of model inference
The number of parameters
Explanation - Interpretability helps clinicians trust and verify AI decisions.
Correct answer is: The extent to which humans can understand why the model made a decision
Q.158 Which imaging modality uses acoustic waves to produce images of the inside of the body?
CT
MRI
Ultrasound
X‑ray
Explanation - Ultrasound transducers emit and receive sound waves, forming images from echoes.
Correct answer is: Ultrasound
Q.159 What is the primary purpose of 'skull stripping' in brain MRI preprocessing?
Remove non‑brain tissues to focus analysis on brain parenchyma
Enhance skull visibility
Reduce image resolution
Add synthetic skull to the image
Explanation - Skull stripping eliminates extracranial tissues that can confuse analysis.
Correct answer is: Remove non‑brain tissues to focus analysis on brain parenchyma
Q.160 Which of the following best describes a 'latent space' in deep learning?
The input image space
A compressed representation of input data within a neural network
The output prediction space
A database of all possible labels
Explanation - Latent spaces capture essential features while reducing dimensionality.
Correct answer is: A compressed representation of input data within a neural network
Q.161 Which of the following is a typical objective of using 'transfer learning' for medical imaging?
Training a model from scratch on limited data
Fine‑tuning a model with pre‑trained weights for improved generalization
Removing the need for labels
Generating synthetic images
Explanation - Transfer learning leverages knowledge learned on large datasets to accelerate learning.
Correct answer is: Fine‑tuning a model with pre‑trained weights for improved generalization
Q.162 Which metric is used to evaluate a segmentation model's boundary accuracy?
Jaccard Index
Hausdorff Distance
Accuracy
Recall
Explanation - Hausdorff distance measures the maximum deviation between predicted and true boundaries.
Correct answer is: Hausdorff Distance
Q.163 What is 'bias‑field correction' used for in MRI?
Correcting for intensity variations across the field of view
Increasing image resolution
Reducing motion artifacts
Enhancing contrast between tissues
Explanation - Bias field correction removes non‑uniformities due to scanner imperfections.
Correct answer is: Correcting for intensity variations across the field of view
Q.164 Which of the following best describes 'focal loss'?
A loss that emphasizes hard examples in imbalanced classification
A loss for regression tasks
A loss for image generation
A loss for classification with balanced classes
Explanation - Focal loss down‑weights easy examples to focus learning on hard, often minority class samples.
Correct answer is: A loss that emphasizes hard examples in imbalanced classification