AI CPD lecture

Transcript

1. ✨ AI in Radiology Dr Daniel Fascia Consultant Musculoskeletal Radiologist Clinical Director of Yorkshire Imaging Collaborative Royal College of Radiologist Informatics Chair What’s out there? What does the law say?
2. In the company of AI legends
3. ✨ All around you The most pivotal change to computing in my life so far
8. Navigate to… Harrogate Hospital
9. Navigating to… Highgate Hospital #&%$£!!
10. Radiology
11. Radiology
12. Radiology Drama Queen?
14. AI is everywhere in Radiology
15. Referral Every stage of imaging with AI Scheduling Scanning Interpretation Report authoring
16. Referral Every stage of imaging with AI Scheduling Scanning Interpretation Report authoring Clinical Decision Support OCR of referrals NLP comprehension Automatic Triage
17. Referral Every stage of imaging with AI Scheduling Scanning Interpretation Report authoring Clinical Decision Support OCR of referrals NLP comprehension Automatic Triage Automatic Protocolling (LLM) Smart booking based on patient factors Interactive advice for appointment (tra ff ic etc)
18. Referral Every stage of imaging with AI Scheduling Scanning Interpretation Report authoring Clinical Decision Support OCR of referrals NLP comprehension Automatic Triage Automatic Protocolling (LLM) Smart booking based on patient factors Interactive advice for appointment (tra ff ic etc) Optimal patient positioning (Camera, Object recognition) Scan acceleration Synthetic upscaling hyperresolution MR Give older scanners new powers
19. Referral Every stage of imaging with AI Scheduling Scanning Interpretation Report authoring Clinical Decision Support OCR of referrals NLP comprehension Automatic Triage Automatic Protocolling (LLM) Smart booking based on patient factors Interactive advice for appointment (tra ff ic etc) Optimal patient positioning (Camera, Object recognition) Scan acceleration Synthetic upscaling hyperresolution MR Give older scanners new powers Pre-reading for safety Simple full interpretation (fracture, OA) Assisted reading (ACL / meniscus, cartilage) Early AI result pre-report
20. Referral Every stage of imaging with AI Scheduling Scanning Interpretation Report authoring Clinical Decision Support OCR of referrals Automatic Triage Automatic Protocolling (LLM) Smart booking based on patient factors Interactive advice for appointment (tra ff ic etc) Optimal patient positioning (Camera, Object recognition) Scan acceleration Synthetic upscaling hyperresolution MR Pre-reading for safety Simple full interpretation (fracture, OA) Co-reading (PE? Intracranial bleed) Early AI result pre-report Fully automated report authoring Simple use cases Regulatory mine f ield
21. What’s out there in MSK?
22. Fracture Detection • Every MSK x-ray analysed (30-60s) • Detects • Adult and paediatric fractures • Effusions • Periprosthetic fractures • Dislocations • Labels images (fracture / no) • Delivers labelled result to PACS in <1-minute
23. Fracture Detection • Every MSK x-ray analysed (30-60s) • Detects • Adult and paediatric fractures • Effusions • Periprosthetic fractures • Labels images (fracture / no) • Delivers labelled result to PACS • Empowers “ED workforce” • Second eyes for others
24. Fracture Detection • Reduce fracture clinic: no fracture • Patient satisfaction: less rollercoaster medicine • Reduced litigation (missed fracture) • £ 35,417,800 (NHS Resolutions) • £6,000 - 14,000 per case • AI solution cost • £20,000 annual Financial Case
25. Knee Osteoarthritis • Works on wide quality range of xray • Pattern of compartmental affect • Severity objectively • Joint space narrowing • Sclerotic changes • Osteophytes • Use cases? • Industrial? for implant selection advice: What do you think? Automated x-ray labelling
26. Hip morphology & OA • Part of overall lower limb package • Morphological assessments • Acetabulum • Proximal femur • Osteoarthritis • Two key vendors to look at • Radiobotics RBfracture, RBKnee, RBHip • Gleamer Bonemetrics, Boneview
27. Osteoporosis - DEXA • Automated reading of DEXA scans • Deeper analysis: microarchitecture helps avoid adjacent degenerative sclerosis • Comparison of skeletal status against statistical data • Automated therapeutic recommendation • Automates reporting of tedium • Can be trained to add advice which is the pinnacle of radiology reporting • Works well: constrained scope
28. Opportunistic vertebral fracture detection • CT scans carried out for many other purposes covering the trunk • “Always review the bones” • Tool to automate bone review • Detect morphology of osteoporosis • Detect vertebral fractures • Measure bone density from CT
29. Bone Age Assessment BoneXpert - quite widespread in UK Automated analysis. Supports pay per use online for low volume situations (smart choice)
30. Why aren’t there more MSK products?
31. Challenging Anatomy
32. Imaging AI in your region Harrogate Leeds Bradford Mid Yorks Airedale Calderdale & Hudders f ield Fracture Detection (RBfracture) Fracture Detection (RBfracture) Fracture Detection (RBfracture) Chest x-ray (Annalise.ai) Chest x-ray (Annalise.ai) Chest x-ray (Behold.ai) Chest x-ray (Annalise.ai) Chest x-ray (Annalise.ai) Chest x-ray (Annalise.ai) Lung cancer (Qure) Breast (Kheiron) CT Brain (Annalise.ai) Stroke (Brainomix)
33. CT Brain • Immediate f irst read of CT brain • Labelled results delivered to PACS • No user interaction needed with tool • Intracranial haemorrhage (types, location, chronicity etc…) • Mass lesion, mass effect • Hydrocephalus • Vasogenic Oedema First read support
34. CT Brain • Immediate f irst read of CT brain • Labelled results delivered to PACS • No user interaction needed with tool • Intracranial haemorrhage (types, location, chronicity etc…) • Mass lesion, mass effect • Hydrocephalus • Vasogenic Oedema • Second eyes for radiology (on-call) • Fast front line opinion (ED, acute med, ICU) First read support
35. CT Brain • 80% of CT brains: nil acute • £73: average cost to outsource overnight • AI tool: £30,000/yr • 8 CT brains per night (quiet place) • Paid for in 52 days Economic model
36. CT Brain • 80% of CT brains: nil acute • £73: average cost to outsource overnight • AI tool: £30,000/yr • 8 CT brains per night (quiet place) • Paid for in 52 days • Cost saving on outsourcing • Immediate results (1-2 mins) • Pressure to report in house next day Economic model
37. CT Brain • Auto analysis of brain CT, angiography and CT perfusion • Segmentation of vessels • Segmentation of suspected • Ischaemic penumbra • Established infarcts Stroke imaging
38. CT Brain • Auto analysis of brain CT, angiography and CT perfusion • Segmentation of vessels • Segmentation of suspected • Ischaemic penumbra • Established infarcts • Empower stroke intervention • Neuroradiologists f inite • Stroke time precious 24/7 Stroke imaging Very compelling use case
39. Chest x-ray
41. REPEAT INDETERMINATE INDETERMINATE
42. REPEAT INDETERMINATE INDETERMINATE Not useful. UI/UX design importance
43. 100% of CXR co-analysed by AI
44. Secondary Captures Deep integration 2025 2023 First AIDF site live 2-months after award
45. Lead time advantage Secondary Captures Deep integration 2025 2023 Yorkshire Imaging Collaborative existing architecture
46. Chest X-ray • 6 x West Yorks & Harrogate Trusts • All chest x-rays (within 1-minute) • Labelled images delivered to PACS • 124 different f indings (“AI features”) • Region level feature parity • Important safety concern Regional Deployment
47. Chest X-ray • Faster supported reporting (maybe) • Faster front line diagnosis • Chest sepsis (faster Rx) • Lung cancer (faster CT/MDTM) • Line and tube safety (NGT) • LIVE throughout YIC Proposed Bene f its
48. AI in Radiology: The Law
49. Regulated by MHRA in the UK AI products are medical devices
50. Regulated by MHRA in the UK AI products are medical devices
51. Regulated by MHRA in the UK AI products are medical devices Most are licensed as Clinical Decision Support Tools Commonly Class IIa IRMER was revised August 2024 Autonomous reporting not allowed
52. Ionising Radiation (Medical Exposure) Regulations • August 2024 revision: Here’s my summary • Human in the loop always required • Reporting must be human issued: responsibility • AI is for clinical decision support, not f inal report • Involve medical physics expert throughout process • Implement audit of the system to assure quality and safety and detect performance drift
53. Filtered by AI discrepancy: here is what we found
56. I can forgive this one: But… why we still need humans in the loop
57. “Skinfold pneumothorax”
58. “Skinfold fracture”
59. Thank you Dr Daniel Fascia Consultant MSK Radiologist Regional Clinical Director of Yorkshire Imaging Collaborative Royal College of Radiologists Informatics Chair @danfascia