Instructions:
● Every question carries 2 marks. Max Marks: 30.
● All the data or value for the questions are given in the question.
● Answer to the point and don't extend the answers.
1. Which is the oldest medical imaging modality? What are the two applications of that
modality? X-ray imaging
2. Which shell electrons can be most easily removed in the given figure for the different shell
levels?
The electrons of the outermost shell, in this figure, N, can most easily be removed.
3. Match the labels for several types of scattering – Absorption, no interaction, Rayleigh,
Compton scattering.
A. No interaction
B. Absorption
C. Rayleigh scattering
D. Compton Scattering
, 4. What is the correct order of ionization in alpha, beta, gamma from highest to lowest?
a. alpha > beta > gamma
b. beta > alpha > gamma
c. gamma > beta > alpha
d. None of the above
5. What are the two advantages of using transfer learning in medical imaging classification?
6. Give two differences between transfer learning and domain adaptation?
7. Which loss is better for training a segmentation model - Jaccard or Dice loss? Why?
8. What are the changes in the model which needs to be done to convert a 2D model to a 3D
model for segmentation? (Mention 2 changes)
9. Give two differences between medical image detection and medical image segmentation?
10. What are the advantages of semi-supervised learning over transfer learning? Give two
advantages.
11. A document can be represented by vector based on number of times certain words occur in
the document. What are two ways of comparing similarity between documents based on these
vectors?
12. Of the two approaches in the above question what is problem with one approach that is
overcome by another approach?
13. Does a trustworthy AI solution warrant the trust in patients and doctors?
14. Guidelines are required to achieve trustworthy AI
a. Yes, trustworthy AI system can be implemented following them
b. No, they only provide a framework, but they are not necessary for the implementation of
trustworthy AI systems
c. Yes, together with mechanisms, trustworthy AI systems can be implemented
d. No, they only provide a set of rules for the evaluation of trustworthy AI
15. Which of the following statements about group convolutional neural networks (G-CNNs) and
data-augmentations are true?
a. Given a finite size dataset, both group convolutional neural networks and data
augmentation are methods that in theory can be used to NN that are guaranteed to
invariant to rotations.
b. It is not to be expected that data augmentations (of whatever kind, e.g. color or scaling
transformations) will improve the performance of G-CNNs.
● Every question carries 2 marks. Max Marks: 30.
● All the data or value for the questions are given in the question.
● Answer to the point and don't extend the answers.
1. Which is the oldest medical imaging modality? What are the two applications of that
modality? X-ray imaging
2. Which shell electrons can be most easily removed in the given figure for the different shell
levels?
The electrons of the outermost shell, in this figure, N, can most easily be removed.
3. Match the labels for several types of scattering – Absorption, no interaction, Rayleigh,
Compton scattering.
A. No interaction
B. Absorption
C. Rayleigh scattering
D. Compton Scattering
, 4. What is the correct order of ionization in alpha, beta, gamma from highest to lowest?
a. alpha > beta > gamma
b. beta > alpha > gamma
c. gamma > beta > alpha
d. None of the above
5. What are the two advantages of using transfer learning in medical imaging classification?
6. Give two differences between transfer learning and domain adaptation?
7. Which loss is better for training a segmentation model - Jaccard or Dice loss? Why?
8. What are the changes in the model which needs to be done to convert a 2D model to a 3D
model for segmentation? (Mention 2 changes)
9. Give two differences between medical image detection and medical image segmentation?
10. What are the advantages of semi-supervised learning over transfer learning? Give two
advantages.
11. A document can be represented by vector based on number of times certain words occur in
the document. What are two ways of comparing similarity between documents based on these
vectors?
12. Of the two approaches in the above question what is problem with one approach that is
overcome by another approach?
13. Does a trustworthy AI solution warrant the trust in patients and doctors?
14. Guidelines are required to achieve trustworthy AI
a. Yes, trustworthy AI system can be implemented following them
b. No, they only provide a framework, but they are not necessary for the implementation of
trustworthy AI systems
c. Yes, together with mechanisms, trustworthy AI systems can be implemented
d. No, they only provide a set of rules for the evaluation of trustworthy AI
15. Which of the following statements about group convolutional neural networks (G-CNNs) and
data-augmentations are true?
a. Given a finite size dataset, both group convolutional neural networks and data
augmentation are methods that in theory can be used to NN that are guaranteed to
invariant to rotations.
b. It is not to be expected that data augmentations (of whatever kind, e.g. color or scaling
transformations) will improve the performance of G-CNNs.
