| Question 1: (#1) What is a framework for transforming diagnostic billing codes into AI-friendly structures? |
| Reference: | Reference: https://www.ncbi.nlm.nih.gov/pmc/articles/PMC6844192/ |
| Choice A: | (a) medical language processing |
| Choice B: | (b) transformer models |
| Choice C: | (c) phenome wide association / phenome disease association studies (PheWAS/PheWAS) |
| Choice D: | (d) disease association studies |
| Question 2: (#2) What were the fundamental innovation that allowed neural networks to be effective on medical images? |
| Reference: | Reference: https://link.springer.com/chapter/10.1007/978-3-319-24574-4_28 |
| Choice A: | (a) convolutional blocks |
| Choice B: | (b) residual connections |
| Choice C: | (c) multi-scale connections |
| Choice D: | (d) downsampling |
| Question 3: 1. It is not possible to train an auto-segmentation model on both contrast enhanced and non-contrast enhanced CT images of the liver. |
| Reference: | Reference: Anderson BM, Lin EY, Cardenas CE, Gress DA, Erwin WD, Odisio BC, Koay EJ, Brock KK. Automated Contouring of Contrast and Noncontrast Computed Tomography Liver Images With Fully Convolutional Networks. Adv Radiat Oncol. 2020 May 16;6(1):100464. doi: 10.1016/j.adro.2020.04.023. PMID: 33490720; PMCID: PMC7807136. |
| Choice A: | True |
| Choice B: | False |
| Question 4: 2. The use of Dice similarity coefficient to measurement accuracy of auto-segmentation |
| Reference: | Reference: Anderson BM, Lin EY, Cardenas CE, Gress DA, Erwin WD, Odisio BC, Koay EJ, Brock KK. Automated Contouring of Contrast and Noncontrast Computed Tomography Liver Images With Fully Convolutional Networks. Adv Radiat Oncol. 2020 May 16;6(1):100464. doi: 10.1016/j.adro.2020.04.023. PMID: 33490720; PMCID: PMC7807136. |
| Choice A: | a. May over or underestimate accuracy compared to clinical acceptability |
| Choice B: | b. Is a perfect surrogate for clinical acceptability |
| Choice C: | c. Is only valid to assess tumor segmentation |
| Choice D: | d. Is only valid to assess organ segmentation |
| Question 5: 3. In a recent study on liver auto-segmentation, the DeepLab v3+ architecture |
| Reference: | Reference: Anderson BM, Lin EY, Cardenas CE, Gress DA, Erwin WD, Odisio BC, Koay EJ, Brock KK. Automated Contouring of Contrast and Noncontrast Computed Tomography Liver Images With Fully Convolutional Networks. Adv Radiat Oncol. 2020 May 16;6(1):100464. doi: 10.1016/j.adro.2020.04.023. PMID: 33490720; PMCID: PMC7807136. |
| Choice A: | a. Performed significantly worse than both the VGG16 and the 3D UNET |
| Choice B: | b. Performed significantly worse than the VGG16 but significantly better than the 3D UNET |
| Choice C: | c. Performed significantly better than the VGG16 but significantly worse than the 3D UNET |
| Choice D: | d. Performed significantly better than both the VGG16 and the 3D UNET |
| Question 6: (#1) When optimizing k-space sampling patterns for MRI acquisition: |
| Reference: | Reference: G. Wang, T. Luo, J-F. Nielsen, D. Noll, J. Fessler: “B-spline parameterized joint optimization of reconstruction and k-space trajectories (BJORK) for accelerated 2D MRI”, 2021. http://arxiv.org/abs/2101.11369 |
| Choice A: | (a) Using a B-spline parameterization automatically ensures that hardware constraints like slew rate limits are enforced. |
| Choice B: | (b) The optimal sampling patterns for different anatomical regions can differ. |
| Choice C: | (c) One must consider the effects of eddy currents in the optimization process. |
| Choice D: | (d) The optimization problem is convex. |
| Question 7: (#2) In terms of MRI reconstruction methods, which of the following statements is true: |
| Reference: | Reference: G. Wang, T. Luo, J-F. Nielsen, D. Noll, J. Fessler: “B-spline parameterized joint optimization of reconstruction and k-space trajectories (BJORK) for accelerated 2D MRI”, 2021. http://arxiv.org/abs/2101.11369 |
| Choice A: | (a) Optimized k-space sampling benefits only deep-learning reconstruction methods. |
| Choice B: | (b) A neural network is mathematically equivalent to a regularization term. |
| Choice C: | (c) An unrolled neural network approach with data consistency steps outperforms an image-domain U-Net approach. |
| Choice D: | (d) Neural network reconstruction methods are faster than compressed sensing methods. |
| Question 8: 1. Which of the following can assist with AI workflow orchestration? |
| Reference: | Reference: Integrating the Healthcare Enterprise: IHE Radiology Technical Framework Supplement for AI Results; Radiology Technical Committee. https://www.ihe.net/uploadedFiles/Documents/Radiology/IHE_RAD_Suppl_AIR.pdf. July 16, 2020; accessed March 24, 2021. |
| Choice A: | a. DICOM |
| Choice B: | b. DOCKER |
| Choice C: | c. FHIR |
| Choice D: | d. IHE |
| Question 9: 2. All of the following are ML inference integration standards except |
| Reference: | Reference: Docker Overview. https://docs.docker.com/get-started/overview/. Accessed March 24, 2021. |
| Choice A: | a. CDE |
| Choice B: | b. DICOM |
| Choice C: | c. DOCKER |
| Choice D: | d. FHIR |
| Question 10: 3. Which of the following standards can be used for AI results reporting? |
| Reference: | Reference: Common Data Elements (CDEs) for Radiology. https://www.radelement.org/. Accessed March 24, 2021.
Clunie DA. DICOM Structured Reporting. https://www.google.com/books/edition/DICOM_Structured_Reporting/EVjOolUJNGUC?hl=en&gbpv=1&pg=PP1&printsec=frontcover. PixelMed Publishing, Bangor, PA, 2000. |
| Choice A: | a. CDE and DICOM |
| Choice B: | b. CDE and DOCKER |
| Choice C: | c. DICOM and DOCKER |
| Choice D: | d. FHIR and IHE |
