Preparing Ultrasound Imaging Data for Artificial Intelligence Tasks: Anonymization, Cropping, and Tagging

D. Pechlivanis, S. Didaskalou, E. Kaldoudi, and G. Drosatos Preparing Ultrasound Imaging Data for Artificial Intelligence Tasks: Anonymization, Cropping, and Tagging presented by Eleni Kaldoudi Athena Research Center, Greece Democritus University of Thrace, Greece ThrombUS+ Horizon Project HEALTHINF 2025| Porto, Portugal | 22 February 2025 more than 350,000 papers related to AI/ML in biomedical applications since 2000 11% of AI/ML papers in biomedicine relate to diagnostic imaging 20% of AI/ML papers in diagnostic imaging relate to ultrasonography 1) "artificial intelligence"[MeSH Terms] OR ("artificial"[All Fields] AND "intelligence"[All Fields]) OR "artificial intelligence"[All Fields] OR ("machine learning"[MeSH Terms] OR ("machine"[All Fields] AND "learning"[All Fields]) OR "machine learning"[All Fields]) 2) "diagnostic imaging"[MeSH Subheading] OR ("diagnostic"[All Fields] AND "imaging"[All Fields]) OR "diagnostic imaging"[All Fields] 3) "ultrasound"[All Fields] OR "ultrasonography"[MeSH Terms] OR "ultrasonography"[All Fields] OR "ultrasonics"[MeSH Terms] OR "ultrasonics"[All Fields] OR "ultrasounds"[All Fields] OR "ultrasound s"[All Fields] #1 | #1 AND #2 | #1 AND #3 PubMed AI/ML papers 2024: 70,458 3 a2024 scoping review identified 86 randomized controlled trials that evaluate artificial intelligence applications in clinical practice finds that the majority ▪involve AI applications in diagnostic imaging ▪report statistically significant healthcare improvement when AI is employed HEALTHINF 2025| Porto, Portugal | 22 February 2025 most important factor for successful application of AI/ML in diagnostic imaging is the training data set facts: −universal use of DICOM standard −rigorous privacy regulations (GDPR, HIPAA) key challenges: −ensure patient anonymity in DICOM data −allow for tagging, labeling DICOM metadata element (0010,0010) value representation patient name examples (0010,0020) value representation patient ID PN LO tag value representation value elementgroup HEALTHINF 2025| Porto, Portugal | 22 February 2025 many free solutions are available for DICOM data de-identification however, de-identification is not always efficient, since: −DICOM implementations vary by vendor and imaging modality (e.g. vendors may introduce their own tags) −very often, personal data is also burned in the image HEALTHINF 2025| Porto, Portugal | 22 February 2025 US-DICOMizer an integrated approach to ultrasound data preparation for AI/ML training sets −DICOM metadata de-identification: customizable, can be inspected −burned-in image cropping: customizable per imaging device −generation of image tags related to training set −batch image processing: customizable workflows Greece Lithuania Germany Italy France Finland Spain USA 18 partners from 8 countries Horizon Europe Project Wearable Continuous Point-of-Care Monitoring, Risk Estimation and Prevention for Deep Vein Thrombosis 9.5 M €| 2024-2027 Coordinator: Ε. Kaldoudi, ATHENA RC, Greece HEALTHINF 2025| Porto, Portugal | 22 February 2025 Deep Vein Thrombosis (DVT) is the clotting of blood in a deep vein of the pelvis or an extremity (usually calf or thigh) ▪affects more than 1,000,000 Americans per year 700,000 Europeans per year ▪1/2 of people with DVT experience a sudden pulmonary embolism ▪about 1/4 of those who have a pulmonary embolism die from it ▪annual health expenditure related to DVT is €8.5 billion in EU in immobile patients after major surgery in cancer patient in pregnancy during flights …. Olaf M et al. Deep Venous Thrombosis. Emerg Med Clin North Am. 2017 Cohen AT et al. Thromb Haemost. 2007 Oct;98(4):756-64 Barco S et al.Thromb Haemost. 2016 Apr;115(4):800-8 HEALTHINF 2025| Porto, Portugal | 22 February 2025 compression ultrasonography is the method of choice for DVT diagnosis when DVT is present, vein does NOT fully collapse artery vein >90% sensitive >95% specific requires examination by an expert HEALTHINF 2025| Porto, Portugal | 22 February 2025 evaluation using DVT ultrasound datasets by 4 different vendors •processing time per frame: −Monochrome2: 19-23 ms −YBR FULL 422: 25-30 ms −RGB: 49-61 ms •full dataset preparation time: 3-6 minutes per ultrasound exam •preliminary user feedback: −high usability and efficiency for medical imaging professionals •extensive usability assessment is currently underway in 5 hospitals (via SUS: System Usability Scale questionnaire) Windows 10 desk-top computer Intel Core i7-7700HQ CPU at 2.8 GHz,16 GB RAM Media Type (resolution) Photometric Interpretation Cropping Area (pixels) AVG Time per Frame (msec) Ultrasound Image (1200x800) [100 runs] Monochrome2 576x432 19 768×576 21 1024×768 23 YBR FULL 422 576x432 25 768×576 27 1024×768 30 RGB 576x432 49 768×576 53 1024×768 61 Ultrasound Multi - frame (1200x800) [10 runs] Monochrome2 576x432 102 768×576 105 1024×768 108 YBR FULL 422 576x432 117 768×576 119 1024×768 124 RGB 576x432 241 768×576 244 1024×768 250 HEALTHINF 2025| Porto, Portugal | 22 February 2025 US-DICOMizer open source, available under MIT License programming Language: Python libraries used: −Pydicom for handling DICOM files −Tkinter for GUI development −SimpleITK, NumPy, Pillow for image processing −Matplotlib for visualization app.thrombus.eu/software Alexandroupoli, Jan 2024 Foggia, Jan 2025 Foggia, Jan 2025 work conducted within the ThrombUS+ Horizon project in collaboration with the MSc in Biomedical Informatics, organized by Democritus University of Thrace and Athena Research Center, Greece https://www.linkedin.com/co mpany/thrombus-eu-project/ https://thrombus.eu/ [email protected] Co-funded by the European Union MSc https://bmi.med.duth.gr/ [email protected]h.gr [email protected] MSc in Biomedical Informatics