AI-Powered Anonymization of Clinical Trial Images

How Synaptiq developed deep learning models to automatically detect and anonymize personally identifiable information in clinical trial images for a pharmaceutical company

Problem:

A large pharmaceutical company is dedicated to advancing clinical research with rigorous privacy and compliance standards. Clinical trial images can contain sensitive visual information such as faces, tattoos, and backgrounds that may identify trial participants. Manual redaction of these images is time-consuming, error-prone, and not scalable. The company needed a robust, automated approach to anonymize images efficiently while preserving data quality for research purposes. As part of their commitment to data security and participant confidentiality, they asked Synaptiq to help them anonymize personally identifiable information (PII) in clinical images from trials using AI.

How can AI find you and make you anonymous?

Aggregation, masking, generalization, perturbation, there are several techniques that can be used to anonymize data. Finding and anonymizing data within a dataset involves several steps to ensure that data is de-identified and cannot be linked back to specific individuals. Achieving true anonymity in a dataset is challenging and requires careful consideration of legal, ethical, and technical factors. But once data has been effectively de-identified and cannot be linked back to specific individuals, you must undergo statistical methods and other techniques to assess the risk of re-identification. In the case of clinical studies, people must be protected, research practices must be ethical, laws must be followed, bias eliminated, and data must be able to be shared. Anonymity facilitates data sharing and promotes collaboration among researchers. When data is anonymized, it can be shared more easily with other researchers, which can lead to increased scientific transparency, reproducibility, and advancement of knowledge. There are many good reasons to be both findable and remain anonymous.

Solution:

Synaptiq worked with the client to develop proof-of-concept models for detecting and masking personally identifiable information (PII) - faces, tattoos, and skin - that have high accuracy using only public data.  The use of public data was essential to get over the initial barrier of external partners working with sensitive data, enabling the client to stand up a detection system internally with human-in-the-loop error correction to achieve significant speedups over manual masking.

Each model served as a foundation for scalable anonymization workflows. The face landmark model provided fine-grained, configurable masking of facial features while the tattoo detection model identified tattoos in most standard clinical image scenarios. The skin segmentation model reliably separated skin from background across a variety of skin tones and lighting conditions. Collectively, these models established a strong baseline for automated PII redaction in clinical trial imagery.

Outcome:

The face landmark detection model extracted 454 facial keypoints, enabled precise masking of critical facial regions (eyes, eyebrows, lips) with configurable polygon masks. The system worked reliably except in rare cases of extreme side profiles.  The tattoo detection model was trained on a public dataset of annotated tattoos, achieving a mean Average Precision (mAP) of 75%.  The skin segmentation model achieved Intersection over Union (IoU) scores of 90%.  Each model was delivered in a Python notebook within a GitHub repository, providing a reusable framework for the engineering teams.

This infrastructure allowed the company to move past expensive, manual masking of images for clinical trials - from several minutes per image to several seconds - to a scalable automated system with human-in-the-loop quality control and continuous improvement.

HUMANKIND OF IMPACT

64,800 patients remain anonymous to offer others a chance to get better. By enabling scalable, automated anonymization of clinical trial images, this work supports both scientific progress and the ethical obligation to protect participant privacy. It allows research institutions to responsibly share visual data for AI development, improving models in healthcare while maintaining trust and confidentiality. This balance of innovation and privacy strengthens public confidence in clinical research, accelerates discovery, and helps ensure that advances in medicine are inclusive, secure, and ethically grounded.

9,000 more people each year will be able to start their new careers in America.

HUMANKIND OF IMPACT