1 / 50

Charles E. Kahn, Jr., MD, MS University of Pennsylvania @ cekahn

Imaging Informatics 2017 Session S42. Charles E. Kahn, Jr., MD, MS University of Pennsylvania @ cekahn. Disclosure. My spouse and I have no relevant relationships with commercial interests to disclose. Learning Objectives.

rigg
Télécharger la présentation

Charles E. Kahn, Jr., MD, MS University of Pennsylvania @ cekahn

An Image/Link below is provided (as is) to download presentation Download Policy: Content on the Website is provided to you AS IS for your information and personal use and may not be sold / licensed / shared on other websites without getting consent from its author. Content is provided to you AS IS for your information and personal use only. Download presentation by click this link. While downloading, if for some reason you are not able to download a presentation, the publisher may have deleted the file from their server. During download, if you can't get a presentation, the file might be deleted by the publisher.

E N D

Presentation Transcript


  1. Imaging Informatics 2017 Session S42 Charles E. Kahn, Jr., MD, MS University of Pennsylvania @cekahn

  2. Disclosure • My spouse and I have no relevant relationships with commercial interests to disclose. AMIA 2017 | amia.org

  3. Learning Objectives • Recognize the year’s most important advances in imaging informatics • Describe new developments in interoperability, NLP, and high-dimensionality data that impact medical imaging • Understand current research in medical image informatics and directions for the future

  4. Outline • Interoperability and communication • Natural language processing • High-dimensionality data • Other articles of note bit.ly/Imaging2017

  5. Interoperability and Communication bit.ly/Imaging2017

  6. Chen JY, et al. RadioGraphics 2017

  7. Standards for Radiology Reports • IHE Management of Radiology Report Templates (MRRT) profile • Uses Web standards to define general structure for radiology reporting templates • DICOM Part 20 • Describes how to transmit reports in HL7 Clinical Document Architecture (CDA) format Chen JY, et al. RadioGraphics 2017

  8. Chen JY, et al. RadioGraphics 2017

  9. Rubin DL. Radiology 2017

  10. CDE Applications • Structured clinical reporting • Research case data forms • Data registries • Quality improvement • Pragmatic trials • Computer-assisted reporting • Structured recommendations

  11. Natural Language Processing (NLP) bit.ly/Imaging2017

  12. Kovacs MD, et al. RadioGraphics 2017

  13. “Correlate” • UMLS Metathesaurus links various source vocabularies • MetaMap detects UMLS concepts within text • PACS knows the “context” of the exam being viewed • Radiologist • Patient ID + demographics • Imaging modality + body area • “Correlate” uses NLP to search unstructured EHR data

  14. Kovacs MD, et al. RadioGraphics 2017

  15. Kovacs MD, et al. RadioGraphics 2017

  16. Kovacs MD, et al. RadioGraphics 2017

  17. Correlate • Leverages NLP  • Searches  EHR for clinically relevant follow-up information regarding a radiologic study • Integrates with PACS • Context: patient, provider, exam • Web-based software • Prospectively searches and parses unstructured clinical reports • Scoring results by likely clinical relevance Kovacs MD, et al. RadioGraphics 2017 AMIA 2017 | amia.org

  18. Methods • Automatically extract clinical findings in radiology reports • Characterize level of change and significance • Radiology-specific information model • Machine learning + rule-based reasoning • 40 chest CT reports • Manual annotation for validation Hassanpour S, et al. J Digit Imaging 2017. AMIA 2017 | amia.org

  19. Hassanpour S, et al. J Digit Imaging 2017. AMIA 2017 | amia.org

  20. Results Hassanpour S, et al. J Digit Imaging 2017. AMIA 2017 | amia.org

  21. New Directions • Extend to multiple radiology reports for each patient • Capture complete imaging history • Temporal patterns across various reports • Non-monotonic reasoning • Pointwisemutual information • Enrich information model • More concept classes, such as urgency Hassanpour S, et al. J Digit Imaging 2017. AMIA 2017 | amia.org

  22. High-Dimensionality Data bit.ly/Imaging2017

  23. Radiomics: noninvasive assessment of molecular and clinical tumor characteristics • Advance clinical decision making by analyzing standard-of-care medical images Grossmann P, et al.  eLife 2017

  24. Methods • Two independent lung-cancer cohorts • 262 North American patients • 89 European patients with lung cancer • Analyze relationships • Imaging features • Immune response • Inflammation • Survival • Immunohistochemical staining AMIA 2017 | amia.org

  25. Grossmann P, et al.  eLife 2017

  26. Grossmann P, et al. eLife2017

  27. Two independent lung cancer cohorts (D1, D2) • Radiomic data (R) • Genomic data (G) • Clinical data (C) • Gene set enrichment analysis • Clinical association • Overall survival (red) • Pathologic histology (purple) • TNM stage (yellow) Grossmann P, et al.  eLife 2017

  28. Results • Imaging features showed predictive value • Intra-tumor heterogeneity  activity of RNA polymerase transcription • AUC = 0.62, p=0.03 • Intensity dispersion  autodegrationpathway of ubiquitin ligase • AUC = 0.69, p<10-4 • Prognostic biomarkers performed best when combined • Radiomic, genetic, and clinical information • CI = 0.73, p<10-9 AMIA 2017 | amia.org

  29. Grossmann P, et al.eLife2017

  30. Methods • Retrospective study • 40 patients (median age, 60 years; age range, 44–71 years) • Prostate MRI with T2-weighted and diffusion-weighted imaging • Subsequent robot-assisted radical prostatectomy • Digital histopathologic analysis (DHA) • Lumen, epithelium, stroma, and epithelial nucleus • Registered with MR images • Correlated on a per-voxel basis • Relationship assessed using linear mixed-effects model and Pearson correlation coefficient Kwak JT, et al. Radiology 2017 AMIA 2017 | amia.org

  31. Results • Significant relationship of MRI + DHA parameters (p < .01) • Lumen density (% area of tissue components) • Differences between benign + malignant regions • Gleason score associated with MRI and DHA parameters (p < .05) • Positively related to high-b-value DW imaging • Negatively related to lumen density Kwak JT, et al. Radiology 2017 AMIA 2017 | amia.org

  32. Kwak JT, et al. Radiology 2017

  33. Kwak JT, et al. Radiology 2017

  34. Other Articles of Note bit.ly/Imaging2017

  35. Chennubhotla C, et al.Yearbook of Medical Informatics 2017

  36. Recommendations • Interoperability • Standards in data collection • Data sharing and validation of imaging tools • Clinician feedback in all phases of R&D • Reproducibility and reusability • Open-source architecture Chennubhotla C, et al.Yearbook of Medical Informatics 2017

  37. Recommendations • Innovation • Offer challenges that simulate the real world • Commercialization • Partner with industry • Education • Translate technology from research domain to clinical utility Chennubhotla C, et al.Yearbook of Medical Informatics 2017

  38. Reduced physicians skills • Overreliance on technology • Text over context • Conclusions may ignore the context • Artificial certainty in medicine • All-or-none classification into categories • AI’s “black box” • Inscrutable models Cabitza F, et al.  JAMA 2017

  39. Reducing the Skills of Physicians • Diagnostic sensitivity • 50 mammogram readers • More discriminating readers were presented with challenging images marked by computer-aided detection • 14% decrease • Diagnostic accuracy • 30 internal medicine residents • ECGs annotated with inaccurate computer-aided diagnoses • Decrease to 48% from 57% Cabitza F, et al.  JAMA 2017 AMIA 2017 | amia.org

  40. Focus on Text and the Demise of Context • ML-DSS works on available data (“text”), lacks context • May lead to partial or misleading interpretations of ML-DSS diagnostic, therapeutic, or prognostic outputs • Could lead to reduced interest in and decreased ability to perform holisticevaluations of patients • Treatment of pneumonia + asthma • 14,199 patients with pneumonia • Risk of death (per ML): pneumonia + asthma < pneumonia without asthma • Patients with h/o asthma who presented with pneumonia usually admitted to ICU Cabitza F, et al.  JAMA 2017 AMIA 2017 | amia.org

  41. Intrinsic Uncertainty in Medicine • “Machine learning–based decision support systems bind empirical data to categorical interpretation” • Reliability and accuracy of machine learning performance is affected by observer variability Cabitza F, et al.  JAMA 2017 AMIA 2017 | amia.org

  42. Machine Learning Black Box https://www.kdnuggets.com/2015/10/top-arxiv-deep-learning-papers-explained.html AMIA 2017 | amia.org

  43. AMIA 2017 | amia.org

  44. ML and Prediction in Medicine • ML approaches identify strong, but theory-free, associations • Foster a stronger appreciation of ML capabilities and limitations • Benchmark against real-world standards of care  • Soften the “trough of disillusionment” • Is AI “smarter” than humans? • Stimulating debate — but largely irrelevant • Better together! • Humans + AI > either alone Chen JH, Asch SM. N Engl J Med 2017 AMIA 2017 | amia.org

  45. Outline • Interoperability and communication • Natural language processing • High-dimensionality data • Other articles of note bit.ly/Imaging2017

  46. Thank you ! ckahn@upenn.edu bit.ly/Imaging2017

More Related