Canonical text (the exact bytes fingerprinted)

TRUVACE RECORD VERSION
record: TRV-2026-1530
version: 1
kind: certified
reason: Certified into the record
timestamp: 2026-10-16T06:53:38.875294Z
status: published
lens: g_space
sector: health
headline: [Design and implementation of a clinical trial image medical record desensitization system based on artificial intelligence technology]
dek: Objective: To designe a clinical trial image medical record desensitization system based on artificial intelligence technology and to conduct preliminary applications, achieving the protection of patient privacy and promoting the safe use of clinical image data in clinical trial projects.
Methods: The clinical trial image record desensitization system based on artificial intelligence technology was first performed to preprocess operations, such as cropping and compression on medical images to initialize the imag…
gain_title: An AI system automatically finds and masks patient identifiers in clinical trial image records, reaching 88.48% average accuracy and 99.40% on serious adverse event documents in sampled checks.
problem_title: (none)
trace_subject: (none)
gain_reading: An AI system automatically finds and masks patient identifiers in clinical trial image records, reaching 88.48% average accuracy and 99.40% on serious adverse event documents in sampled checks.
gain_evidence: The average accuracy rate of medical document desensitization reached 88.48% | an accuracy rate of up to 99.40% for serious adverse event medical documents
problem_reading: (none)
problem_evidence: (none)
quick_read: Researchers designed a system that uses image recognition, natural language processing and convolutional neural networks to classify medical documents in clinical trial image records. It finds personal identifiers such as names, ID numbers, medical record numbers, phone numbers and birthplaces, then applies pixel-level masking or blurring.

After the system processed 100,000 image records, 9,000 samples were randomly chosen to check the automated desensitization. The authors say this supports privacy protection and safer sharing of trial image data. Accuracy varied by document type, and the abstract reports the system is paired with a manual online step.
limitation: Average accuracy across document types was 88.48%, so some documents were not correctly desensitized, and the system is described as working together with a manual online auxiliary system.
tag: Evidence-backed gain
key_points: The pipeline first preprocesses images by cropping and compressing them before classification and text recognition. | Targeted identifiers include patient name, identification number, medical record number, phone number and birthplace. | Verification used 9,000 randomly selected samples from 100,000 processed image records. | Average accuracy was 88.48%, rising to 99.40% for serious adverse event documents. | A canvas-based manual online desensitization system assists the automated step.
rundown: The system starts by preprocessing images, then classifies document content through labeling, feature extraction and text recognition. Sensitive points it finds are masked or blurred at pixel level.

For validation, 100,000 image records were desensitized and 9,000 samples from the processed document types were randomly selected for automated effect verification. Average accuracy across document types was 88.48%.

Performance was strongest for serious adverse event medical documents, at up to 99.40%. The authors conclude the tool, combined with the manual online auxiliary system, can be used in production environments for safe sharing of trial image records.
sources:
- peer_reviewed | Beijing da xue xue bao. Yi xue ban = Journal of Peking University. Health sciences | https://pubmed.ncbi.nlm.nih.gov/42880929/ | 2026-10-18
prev: 0000000000000000000000000000000000000000000000000000000000000000

sha256

7899506035ff298c86de907b5e8d93183a9d3731e91f78ae44ecde5843a31884

previous

0000000000000000000000000000000000000000000000000000000000000000