| Question 1: For a K-class classification problem, the last layer of the neural network often maps a K-dimensional vector from the next-to-last layer into another K-dimensional vector as the final output. The elements of the final output sum up to one, and each element is in the range (0,1). What is this mapping function called?
|
| Reference: | Lawrence S, Giles CL, Tsoi AC, Back AD. Face recognition: A convolutional neural-network approach. IEEE Transactions on Neural Networks. 1997 Jan;8(1):98-113.
|
| Choice A: | The activation function. |
| Choice B: | The softmax function. |
| Choice C: | Receptive field. |
| Choice D: | Pooling. |
| Question 2: Which of the following is true for “pooling” in a convolutional neural network (CNN)?
|
| Reference: | LeCun Y, Bengio Y, Hinton G. Deep learning. Nature. 2015 May 28;521(7553):436-44.
|
| Choice A: | Pooling combines information from a patch of a CNN layer to be used as input into the next layer. |
| Choice B: | Pooling helps translation invariance of the CNN. |
| Choice C: | Pooling reduces the image size from one layer to the next. |
| Choice D: | All of the above. |
| Question 3: Which of the following is NOT aimed at reducing overfitting of a convolutional neural network?
|
| Reference: | Krizhevsky A, Sutskever I, Hinton GE. Imagenet classification with deep convolutional neural networks. NIPS'12 Proceedings of the 25th International Conference on Neural Information Processing Systems, 2012 (pp. 1097-1105).
|
| Choice A: | Backpropagation. |
| Choice B: | Data augmentation. |
| Choice C: | Dropout. |
| Choice D: | Early stopping. |
| Question 4: What is the main challenge in building robust deep learning CNN (DL-CNN) models in the field of medical imaging?
|
| Reference: | Srivastava, N., Hinton, G.E., Krizhevsky, A., Sutskever, I. and Salakhutdinov, R., Dropout: a simple way to prevent neural networks from overfitting, Journal of Machine Learning Research, 2014, 15(1), pp.1929-1958.
|
| Choice A: | Lack of freely available DL-CNN software packages. |
| Choice B: | Lack of sufficiently large memory to hold the DL-CNN structure. |
| Choice C: | Lack of a large enough data set to have independent training, validation, and test partitions. |
| Choice D: | To reach high accuracy on the training data set. |
| Question 5: The transfer learning technique for the deep learning networks is characterized best by: |
| Reference: | Yosinski J., Clune J., Bengio Y., Lipson H., How transferable are features in deep neural networks?, Advances in Neural Information Processing Systems 27, pp. 3320-3328, Dec. 2014.
|
| Choice A: | Fast training of the entire deep learning CNN structure. |
| Choice B: | Fixing the parameters of part of a deep learning CNN already trained with a large data set and retraining the rest of the structure with a new data set relevant to the specific application of interest. |
| Choice C: | Large learning rate. |
| Choice D: | Backpropagation. |
| Question 6: Convolutional neural networks are multi-layer artificial neural networks, which can be employed in various (sub)tasks along the image analysis pipeline. Which of the following can be enabled with a CNN?
|
| Reference: | Zhang, W. et al. Computerized detection of clustered microcalcifications in digital mammograms using a shift-invariant artificial neural network. Med. Phys. 21, 517–524 (1994).
•Huynh, B. Q., Li, H. & Giger, M. L. Digital mammographic tumor classification using transfer learning from deep convolutional neural networks. J Med Imaging (Bellingham) 3, 034501 (2016).
•D. Ravi, C. Wong, F. Deligianni, M. Berthelot, J. Andreu-Perez, B. Lo, and G. Z. Yang, "Deep Learning for Health Informatics," IEEE Journal of Biomedical and Health Informatics 21, 4-21, (2017).
|
| Choice A: | Image filtering. |
| Choice B: | Tumor segmentation. |
| Choice C: | Disease state classification. |
| Choice D: | All of the above. |
| Question 7: Given the large amount of image data needed to train a deep CNN, one can augment the training effort by which of the following methods?
|
| Reference: | Yosinski, J., Clune, J., Bengio, Y., Lipson, H, “How transferable are features in deep neural networks?,” in Advances in Neural Information Processing Systems 27 (eds. Ghahramani, Z., Welling, M., Cortes, C., Lawrence, N. D. & Weinberger, K. Q.) 3320–3328 (Curran Associates, Inc., 2014).
•N. Tajbakhsh, J. Y. Shin, S. R. Gurudu, R. T. Hurst, C. B. Kendall, M. B. Gotway, and J. M. Liang, "Convolutional Neural Networks for Medical Image Analysis: Full Training or Fine Tuning?," IEEE Transactions on Medical Imaging 35, 1299-1312, (2016).
|
| Choice A: | Use pre-trained CNNs, which have been trained with other types of images. |
| Choice B: | Manipulate the input images in order to “generate” more data with which to train. |
| Choice C: | Change the classification task. |
| Choice D: | A and B. |
| Choice E: | All of the above. |
| Question 8: Colitis can be caused by: |
| Reference: | Thoeni RF, Cello JP. CT Imaging of Colitis. Radiology. 2006; 240(3):623-38.
|
| Choice A: | Inflammatory bowel disease. |
| Choice B: | Inflammation in immunocompromised patients. |
| Choice C: | Hospital acquired infection. |
| Choice D: | All of the above. |
| Question 9: Which of the following are least frequently used for training deep learning networks for applications in radiology: |
| Reference: | Shin H-C, Lu L, Kim L, Seff A, Yao J, Summers RM. Interleaved Text/Image Deep Mining on a Very Large-Scale Radiology Database. The IEEE Conference on Computer Vision and Pattern Recognition (CVPR)2015; 1090-9.
|
| Choice A: | Crowdsourcing. |
| Choice B: | Transfer learning. |
| Choice C: | Manual annotation by experts. |
| Choice D: | Natural language processing of radiology reports. |
| Question 10: Which of the following is not a generic platform or software library for performing deep learning research: |
| Reference: | Erickson BJ, Korfiatis P, Akkus Z, Kline TL. Machine Learning for Medical Imaging. Radiographics. 2017; 37(2):505-15.
|
| Choice A: | AlexNet. |
| Choice B: | TensorFlow. |
| Choice C: | Caffe. |
| Choice D: | Theano. |
