| Question 1: Which of the following data strategies are considered good practice for improving the soundness and applicability of deep learning methods? |
| Reference: | David A. Bluemke, Linda Moy, Miriam A. Bredella, Birgit B. Ertl-Wagner, Kathryn J. Fowler, Vicky J. Goh, Elkan F. Halpern, Christopher P. Hess, Mark L. Schiebler, and Clifford R. Weiss, Assessing Radiology Research on Artificial Intelligence: A Brief Guide for Authors, Reviewers, and Readers—From the Radiology Editorial Board, Radiology 2020 294:3, 487-489 |
| Choice A: | (a) Carefully design all three image sets (training, tuning and testing sets) to be independent without overlap. |
| Choice B: | (b) Use an external test set (eg, from another institution) for final statistical reporting |
| Choice C: | (c) Use multivendor images for each phase of the DL evaluation (training, tuning and testing sets) |
| Choice D: | (d) All of the above |
| Question 2: Deep learning-based image reconstruction/denoising methods are data-driven, nonlinear and do not have to rely on imaging physics. To evaluate the performance of a DL reconstruction/denoising network, |
| Reference: | Wang, G., Ye, J.C. & De Man, B. Deep learning for tomographic image reconstruction. Nat Mach Intell 2, 737–748 (2020). https://doi.org/10.1038/s42256-020-00273-z |
| Choice A: | Classic image quality metrics such as mean squared errors (MSE) and structure similarity index measure (SSIM) are of great value. |
| Choice B: | Classic image quality metrics such as mean squared errors (MSE) and structure similarity index measure (SSIM) are of reduced value. |
| Choice C: | Task-based performance metrics become the most relevant. |
| Choice D: | a and c |
| Choice E: | b and c |
| Question 3: Linear regression, principal component analysis, random forests, and neural networks are all machine learning methods employed for dose prediction in automated treatment planning? |
| Reference: | Babier, et al. [dio: 10.1002/mp.13896] |
| Choice A: | True |
| Choice B: | False |
| Question 4: Generalized Adversarial Networks (GANs) cannot be used for dose prediction |
| Reference: | Babier, et al. [doi: 10.1002/mp.13896] |
| Choice A: | True |
| Choice B: | False |
| Question 5: Dice and Gamma metrics can be used to validate output from dose prediction methods |
| Reference: | McIntosh, et al. [doi: 10.1088/1361-6560/62/2/415] |
| Choice A: | True |
| Choice B: | False |
| Question 6: What are the limitations of applying fully convolutional networks for prostate cancer prediction using multiparametric MRI? |
| Reference: | He, K., Gkioxari, G., Dollár, P., & Girshick, R. (2017). Mask r-cnn. In Proceedings of the IEEE international conference on computer vision (pp. 2961-2969).
Wang, X., Girshick, R., Gupta, A., & He, K. (2018). Non-local neural networks. In Proceedings of the IEEE conference on computer vision and pattern recognition (pp. 7794-7803). |
| Choice A: | Predictions don’t separate individual lesions. |
| Choice B: | Convolutional operation only processes one the local neighborhood at a time. |
| Choice C: | The misalignment between MR imaging modalities can compromise segmentation accuracy. |
| Choice D: | All of the above. |
| Question 7: What is the main difference between DQN and traditional supervised machine learning algorithms? |
| Reference: | Mnih, V., Kavukcuoglu, K., Silver, D., Rusu, A.A., Veness, J., Bellemare, M.G., Graves, A., Riedmiller, M., Fidjeland, A.K., Ostrovski, G. and Petersen, S., 2015. Human-level control through deep reinforcement learning. nature, 518(7540), pp.529-533. |
| Choice A: | DQN does not need training data. |
| Choice B: | DQN generates a large amount of training data by interacting with the environment. |
| Choice C: | DQN takes less time to train than supervised machine learning algorithms. |
| Choice D: | DQN does not need independent data set for testing. |
| Question 8: Why interpretability of AI is important for medical physics? |
| Reference: | Xing L, Giger ML, Min JK (editors): Artificial Intelligence in Medicine: Technical Basis and Clinical Applications. Academic Press, St. Louis, MO, 2020. |
| Choice A: | It can aid in trust of the model. |
| Choice B: | It can help in safe use of the AI. |
| Choice C: | It can speed up the model prediction. |
| Choice D: | A and B |
| Choice E: | All of the above. |
| Question 9: What are the potential pitfalls of a data-driven model? |
| Reference: | Xing L, Giger ML, Min JK (editors): Artificial Intelligence in Medicine: Technical Basis and Clinical Applications. Academic Press, St. Louis, MO, 2020.
Shen L, Zhao W, Xing L: Patient-specific reconstruction of volumetric computed tomography images from a single projection view via deep learning. Nature Biomedical Engineering, 3: 880-8, 2019. |
| Choice A: | It requires a large amount of annotated data to build the model. |
| Choice B: | The model may be less trustworthy. |
| Choice C: | The results may be less interpretable. |
| Choice D: | All of the above. |
| Question 10: What are the potential strategies to improve the interpretability of AI models? |
| Reference: | Seo H, Bassenne M, Xing L: Closing the Gap between Deep Neural Network Modeling and Biomedical Decision-Making Metrics in Segmentation via Adaptive Loss Functions. IEEE Transactions on Medical Imaging, 40: 585-93, 2021.
Islam MT, Xing L: A data-driven dimensionality-reduction algorithm for the exploration of patterns in biomedical data. Nature Biomedical Engineering, 2020. doi: 10.1038/s41551-020-00635-3. |
| Choice A: | Incorporation of prior knowledge. |
| Choice B: | Better understanding of feature space data. |
| Choice C: | Analysis of model behavior under some known perturbations. |
| Choice D: | All of the above. |
