OpenTalks.AI - Сергей Кастрюлин, Ускорение МРТ с помощью нейронных сетей

Transcript

1. © Philips - Confidential Nicola Pezzotti1, Elwin de Weerdt2, Sahar Yousefi3, Mohamed S. Elmahdy3, Jeroen van Gemert2, Christophe Schülke1, Mariya Doneva1, Tim Nielsen1, Sergey Kastryulin1, Boudewijn P.F. Lelieveldt3, Matthias J.P. van Osch3, Marius Staring3 1 Philips Research 2 Philips Healthcare 3 Leiden University Medical Center FastMRI with Adaptive Intelligence Best performer in the FastMRI challenge (NeurIPS 2019) multi-coil 4x and 8x
2. © Philips - Confidential MRIs give a wealth of information… … but are slow (~20 min up to an hour)
3. © Philips - Confidential data information From scanner to report
4. © Philips - Confidential data information From scanner to report
5. © Philips - Confidential The more data desire, the more data required Image reconstruction Raw data Image MRI scanner
6. © Philips - Confidential Why so slow? Gathering data takes time
7. © Philips - Confidential Why so slow? Gathering data takes time
8. © Philips - Confidential Why so slow? Gathering data takes time
9. © Philips - Confidential Why so slow? Gathering data takes time
10. © Philips - Confidential Why so slow? Gathering data takes time
11. © Philips - Confidential The more data desire, the more data required Image reconstruction Raw data Image
12. © Philips - Confidential Learned reconstruction with a neural network Raw data Image
13. © Philips - Confidential Three different challenges Single-Coil 4x Easier to experiment, simulated single coil measurement. Not used clinilcally. Multi-Coil 4x Clinically relevant, multiple coil measurements. ~2x faster than parallel imaging. Multi-Coil 8x Clinically relevant, multiple coil measurements. ~4x faster than parallel imaging.
14. © Philips - Confidential Realistic scenario vs ease of experimentation Almost all MRI scanners use technique called “parallel imaging” that uses multiple Receiver coils in order to capture information from different spatial locations. This can speed up scanning process by ~2x.
15. © Philips - Confidential Evaluation Stage 1 SSIM (Structural Similarity Index Metric) measures the perceived change in structural information between two images. As with other quantitative metrics, may not capture qualitative differences perceived by radiologists. Stage 2 Top 4 entries as scored by SSIM are ranked by 7 radiologists. Winners determined by the best average radiologist ranking.
16. © Philips - Confidential Baseline provided by organizers IFFT U-Net
17. © Philips - Confidential Adaptive CS-Network A Deep Learning solution that integrates Compressed-Sensing theory
18. © Philips - Confidential Compressed Sensing for MR reconstruction 18 Iteration n Iteration n+1 Initial Image Intermediate Final Image F(): Sparsifying transform F-1(): Invert Sparsification Thresholding Function Thresholding Function Wavelet decomposition Wavelet decomposition
19. © Philips - Confidential Adaptive CS-Network Combination of Compressed-Sensing theory and Deep-Learning Iterative AI-Based Compressed-Sensing approach + + + + + Iterative – Multiscale – MR Physics •
20. © Philips - Confidential Adaptive CS-Network " . $ " . " . Filtered-UNet • UNet-like design • Each feature map is filtered • Only basic DL operations to retain control • Convolution 2D without bias term • Leaky ReLU • Max pooling • Interpolation 2D • 2.5D Convolutions • Loss applied to the central slice only " . $ " . " .
21. © Philips - Confidential Adaptive CS-Network ' ( ) "*( " ' . ( . " . " . $ " . " . " "*( Deep-Learning Based
22. © Philips - Confidential Adaptive CS-Network ' ( ) "*( " ' . ( . " . -," / -0 - (" , ) / (" ) -0 (" ) " . $ " . " . " "*( + Deep-Learning Based MR-Specific Knowledge
23. © Philips - Confidential Loss function definition Multiscale-SSIM We adopted a mixed approach Loss =0.86 * MSSSIM + 0.14 * L1 • Better results reported in literature • Higher SSIM metric • Less artifacts in the image
24. © Philips - Confidential Loss function definition Panels of diverse reconstructions were shown to radiologists and clinical scientists
25. © Philips - Confidential Winners Single-coil 4x University of Amsterdam Multi-coil 4x United Imaging Intelligence/John Hopkins University Philips Multi-coil 8x Philips
26. © Philips - Confidential Zero Filled Reconstruction Ground-truth Volume: 1000108 – Effective acceleration: 3.64x – SSIM: 0.9885
27. © Philips - Confidential Zero Filled Reconstruction Ground-truth Volume: 1002340 – Effective acceleration: 9.68x – SSIM: 0.8254
28. © Philips - Confidential Conclusion • Improvement over compressed-sensing is within reach • Importance to involve Radiologist and Clinical Scientist • To design loss functions • To evaluate results • Importance of combining prior-knowledge in Deep- Learning solutions • Compressed sensing • MR Physics Paper from NeurIPS’19: https://arxiv.org/abs/1912.12259
29. © Philips - Confidential The Team
30. © Philips - Confidential Nicola Pezzotti1, Elwin de Weerdt2, Sahar Yousefi3, Mohamed S. Elmahdy3, Jeroen van Gemert2, Christophe Schülke1, Mariya Doneva1, Tim Nielsen1, Sergey Kastryulin1, Boudewijn P.F. Lelieveldt3, Matthias J.P. van Osch3, Marius Staring3 1 Philips Research 2 Philips Healthcare 3 Leiden University Medical Center Q&A