More Related Content
Similar to AI-powered Medical Imaging Analysis for Precision Medicine (20)
![[Review] High-performance medicine: the convergence of human and artificial i...](https://cdn.slidesharecdn.com/ss_thumbnails/reviewhigh-performancemedicine-190728142338-thumbnail.jpg?width=560&fit=bounds)
![[Taiwan AI Academy] Machine learning and deep learning application examples](https://cdn.slidesharecdn.com/ss_thumbnails/aiamgrsharing170127-180128082739-thumbnail.jpg?width=560&fit=bounds)
![[Python - Deep Learning] Data generator](https://cdn.slidesharecdn.com/ss_thumbnails/datageneratortopublic-180111155736-thumbnail.jpg?width=560&fit=bounds)
AI-powered Medical Imaging Analysis for Precision Medicine
- 1. AI-Powered Medical Imaging Analysis for Precision Medicine Sean Yu, VP of Data & Machine Learning 2021/05/01
- 2. aetherAI Asia’s Leading Medical Image AI Company • Founded in Oct 2015 • Digital pathology and AI-Powered Medical Imaging • Largest share of the digital pathology market in Taiwan • 60+ Employees (MDs, Data scientists, Web engineers, Systems engineers, MBAs) • 5+ Main products • 10+ Ongoing medical image AI projects • 7 Journal papers in the past two years
- 3. Research and Business Partnerships
- 4. aetherSlide Human-centered & AI-ready Platform for Pathologists A powerful cloud-based platform integrates pathologists’ knowledge and advanced AI techniques to achieve precision medicine
- 5. Cancer Identification - AI as a screening tool
- 6. Extreme Resolution of Whole-Slide Images ImageNet: 224 x 224 Chest X-ray: 2,000 x 2,000 400x WSI: 200,000 x 100,000
- 7. Extreme Resolution of Whole-Slide Images ImageNet: 224 x 224 Chest X-ray: 2,000 x 2,000 400x WSI: 20,000 megapixel iPhone 12 Pro Max 12 megapixel Google Pixel 5 12.2 megapixel
- 8. Developing Cancer Classifier - Common Workaround：Patch-based Method
- 9. ● Laborious human annotations ● Inconsistency between experts Developing Cancer Classifier - Common Workaround：Patch-based Method
- 10. ● Laborious human annotations ● Inconsistency between experts Developing Cancer Classifier - Common Workaround：Patch-based Method
- 11. ● Trapped in Local minimum ● Inefficient information utilization ● Atypical patterns? Developing Cancer Classifier - Advanced Workaround：Multiple Instance Learning Campanella et al. (2019)
- 12. ● Let a CNN consumes whole-slide images DIRECTLY. ● Slide-level ground-truths can be automatically acquired from medical records, scalable to all accessible data in medical centers. (~1,000 slides/day v.s. 8 annotated-slides/day) Developing Cancer Classifier - Our Method：Whole-Slide with Huge Model Support
- 13. Nasopharyngeal Carcinoma Classification Whole-Slide model VS. Patch-based model Benign or NPC ? Two-stage workflow for cancer screening
- 14. Nasopharyngeal Carcinoma Classification Whole-Slide model VS. Patch-based model Patch-based model Whole-Slide model Input Size Effect Training Size Effect ~ 1,000 slides
- 15. Conducting Whole-Slide Training - Operation optimization for practical use From 1 month to 3 days
- 16. Conducting Whole-Slide Training - Operation optimization for practical use From 1 month to 3 days
- 17. Lung Adenocarcinoma/Squamous-cell Carcinoma Classification Whole-Slide model VS. MIL model ~ 10,000 slides
- 18. Application - Cancer Screening on aetherSlide • Fast navigation to hot- spots • Cancer purity of – whole-slide images – within selected region Select Candidate region for NGS Micro metastasis localization
- 19. IHC Module - AI as a quantification tool Use Case • Ki67 / ER / PR • PD-L1 / Her2 Special Use Case • H3K27 • Ki67&CD68 double staining Ki67 H3K27me3 PD-L1 Her2
- 20. on 3 high-power-fields (HPFs)? or whole-slide image? sample 100 cells? 500 cells? or more cells? IHC Response Estimation on Whole-slide Images - Human evaluation can be subjective & discrete
- 21. IHC Response Estimation on Whole-slide Images - Human evaluation can be subjective & discrete Allred scoring system for ER/PR Ki-67, WHO (2017) Credit: NigerJSurg (2018)
- 22. IHC Response Estimation on Whole-slide Images - Human evaluation can be subjective & discrete Agilent Dako, PD-L1 IHC 22C3 Scoring Guideline Roche PD-L1, SP142
- 23. IHC Response Estimation on Whole-slide Images
- 24. IHC Response Estimation on Whole-slide Images Positive Rate ~ 53%
- 25. Automated Ki67 quantification - Performance comparison 10-20 min ?
- 26. Special case of automated Ki67 quantification - Ki67/CD68 double staining Ki67- / CD68+ Ki67+ / CD68+ Ki67+ / CD68- Average correlation: 0.882
- 27. Special case of automated nuclei quantification - H3K27me3 response evaluation
- 30. • Can consuming more data • Can have objective criteria, which is consistent across • People • Time • Space • Can identify more features/parameters correlated to • Clinical indexes • Prognosis • Treatment response In the future … With AI-powered Medical Image Analysis System NTUH: 1,250 slides / day CGMH: 3,800 slides / day
- 31. • Can consuming more data • Can have objective criteria, which is consistent across • People • Time • Space • Can identify more features/parameters correlated to • Clinical indexes • Prognosis • Treatment response In the future … With AI-powered Medical Image Analysis System Karimi et al. (2019) Mercan et al. (2020)
- 32. • Can consuming more data • Can have objective criteria, which is consistent across • People • Time • Space • Can identify more features/parameters correlated to • Clinical indexes • Prognosis • Treatment response In the future … With AI-powered Medical Image Analysis System Ole-Johan Skrede et. al. (2020) Echle et al. (2021)
- 33. From precision diagnosis to precision medicine
- 34. Prevision the Future of AI