C22 Past papers

Question 1

Q: What is the motion constraint equation in optical flow and what does it assume?

Answer 1

A: dI/dx vx plus dI/dy vy plus dI/dt equals zero assumes small intensity changes over time and brightness constancy of pixels

Question 2

Q: In Horn Schunck optical flow why is there a regularization term and what is its role?

Answer 2

A: The term alpha times gradient norm squared of velocity imposes smoothness on the flow field preventing abrupt changes and solving the aperture problem

Question 3

Q: What is the difference between T1 weighted and T2 weighted MRI images?

Answer 3

A: T1 weighted uses short TR and TE fat bright fluid dark whereas T2 weighted uses long TR and TE fluid bright fat dark

Question 4

Q: Write the Bloch equation and explain the significance of each term?

Answer 4

A: dM/dt equals gamma M cross B minus R times M minus M0 the first term is precession the second represents T1 and T2 relaxation toward equilibrium

Question 5

Q: In medical image registration why is mutual information commonly used for multimodal alignment?

Answer 5

A: MI captures statistical dependence between intensity distributions and does not require similar intensity scales thus robust across modalities

Question 6

Q: Define normalized cross correlation and when it is typically used?

Answer 6

A: NCC equals sum of zero mean products divided by product of standard deviations used for single modality registration or template matching

Question 7

Q: What is the main difference between rigid and affine transformations?

Answer 7

A: Rigid preserves shape with rotation and translation only affine also allows scaling and shear

Question 8

Q: Describe the shape of the Haar wavelet in one dimension?

Answer 8

A: Piecewise constant plus one for first half minus one for second half zero outside

Question 9

Q: What does anisotropic diffusion filtering do to an image?

Answer 9

A: It smooths within homogeneous regions while preserving edges by reducing diffusion across large gradients

Question 10

Q: In active contours what do the internal and image energy terms represent?

Answer 10

A: Internal energy enforces curve smoothness while image energy attracts the curve to edges or desired features

Question 11

Q: What is the partial volume effect?

Answer 11

A: When one voxel contains more than one tissue type producing mixed intensities and blurred boundaries

Question 12

Q: How would you compute the Jacobian of a deformation field in registration and why is it important?

Answer 12

A: Determine the determinant of the spatial derivative matrix of the transform it indicates local volume change

Question 13

Q: State the formula for the Dice similarity coefficient?

Answer 13

A: Two times the intersection volume divided by the sum of the individual volumes of sets A and B

Question 14

Q: What is a Bland Altman plot used for?

Answer 14

A: To compare two measurement methods by plotting their mean against their difference revealing bias and agreement limits

Question 15

Q: Give the steps of a two dimensional Procrustes alignment of point sets?

Answer 15

A: Center the points scale to equal norm then compute rotation via SVD to minimize squared distances

Question 16

Q: Define sensitivity and specificity?

Answer 16

A: Sensitivity equals TP over TP plus FN specificity equals TN over TN plus FP

Question 17

Q: Explain positive predictive value also called precision?

Answer 17

A: PPV equals TP over TP plus FP the proportion of positive predictions that are correct

Question 18

Q: What is the Receiver Operating Characteristic curve?

Answer 18

A: Plot of true positive rate versus false positive rate across thresholds assessing classifier performance

Question 19

Q: How is the C statistic or AUC computed from an ROC curve?

Answer 19

A: It is the numerical area under the ROC curve with one for perfect classifier and half for random

Question 20

Q: What is an advantage of using singular value decomposition for analyzing biomedical signals?

Answer 20

A: SVD reveals principal components reduces dimensionality and denoises data while preserving main structure

Question 21

Q: In Lucas Kanade optical flow how is velocity solved?

Answer 21

A: Using least squares over a patch to solve the two by two system based on local spatial and temporal gradients

Question 22

Q: Why might an elastic registration model fail for large local deformations?

Answer 22

A: Elastic models are too stiff and cannot accommodate large non uniform deformations that a fluid model handles better

Question 23

Q: In MRI slice selection how do you select slice thickness?

Answer 23

A: Slice thickness equals excitation bandwidth divided by gamma times gradient amplitude

Question 24

Q: Explain the difference between T2 star and T2?

Answer 24

A: T2 star includes field inhomogeneity effects making it shorter than pure spin spin relaxation time T2

Question 25

Q: In logistic regression what is the interpretation of a coefficient beta i?

Answer 25

A: It is the change in log odds of the outcome for one unit increase in feature xi holding others constant

Question 26

Q: What is the Fourier transforms role in MRI reconstruction?

Answer 26

A: Inverse Fourier transform converts k space data to spatial domain image

Question 27

Q: In wavelet denoising how are coefficients thresholded?

Answer 27

A: Detail coefficients below a chosen threshold are set to zero or reduced soft thresholding to suppress noise while retaining true signal

Question 28

Q: Why can graph cut segmentation be better than simple thresholding?

Answer 28

A: Graph cuts incorporate both data fidelity and spatial smoothness producing coherent segments with global energy minimization

Question 29

Q: Name two ways to handle imbalanced data when training a classifier on medical data?

Answer 29

A: Resample by oversampling minority or undersampling majority or use class weighting or synthetic examples

Question 30

Q: Why might we prefer a kernel approach to a linear SVM?

Answer 30

A: Kernels implicitly map data to high dimensional space enabling non linear decision boundaries without explicit transformation

Question 31

Q: In non rigid registration what is the role of a regularization term on the deformation field?

Answer 31

A: It penalizes non smooth deformations leading to physically plausible transformations and preventing overfitting

Question 32

Q: Provide a single sentence definition of Markov Random Field segmentation?

Answer 32

A: An MRF assigns labels by minimizing a global energy that combines local data likelihood with neighborhood smoothness priors

Question 33

Q: Which transformation model typically has more parameters thin plate spline or affine?

Answer 33

A: Thin plate spline because it includes many local warping coefficients whereas affine has only a few global parameters

Question 34

Q: What evaluation metric would you use to measure alignment of landmark points after registration?

Answer 34

A: The average or root mean square Euclidean distance between corresponding landmarks

Question 35

Q: In wavelet analysis which coefficients capture high frequency content approximation or detail?

Answer 35

A: Detail coefficients capture high frequency components such as edges and transients

Question 36

Q: Why use shrinkage LDA over ordinary LDA in brain computer interface applications?

Answer 36

A: Shrinkage regularizes covariance estimation improving classification when data dimensionality is high and samples are few

Question 37

Q: What is the F1 score and why is it useful for imbalanced datasets?

Answer 37

A: The harmonic mean of precision and recall balancing false positives and false negatives thus informative when classes are uneven

Question 38

Q: Give two advantages of computerized guidelines over standard textual clinical guidelines?

Answer 38

A: They can provide automatic decision support and integrate individual patient data at point of care

Question 39

Q: What is a Kaplan Meier curve and where is it used?

Answer 39

A: A step function estimating survival probability over time applied in time to event analysis such as patient survival studies

Question 40

Q: In clinical workflow modeling name one standard notation and an advantage of mapping processes?

Answer 40

A: BPMN provides a visual representation improving clarity of tasks decisions and responsibilities enabling optimization