Face and ear obfuscation

Customer: Biospective (Montreal, Canada)

Summary: The use of image obfuscation and anonymization is a common practice for privacy protection. However, many computer vision approaches overlook privacy by assuming access to the original, unobfuscated images. In the medical field, even when anonymizing DICOM tags, researchers and clinicians often do not anonymize patients' faces. With the advancement of artificial intelligence, it is becoming easier to identify individuals based on their data. In addition, it is possible to reconstruct medical series in a full 3D view, allowing heads to be viewed from different angles and extracting additional patient characteristics such as gender, age, race, and hairstyle. To address this problem, we developed a model that searches for faces and ears in CT/MRI data and anonymizes them using a specific image processing algorithm, such as mosaicing or blurring. One of the key features of the model is that it does not slice any part of the brain, leaving the entire brain untouched, while still detecting and anonymizing the face and both ears. For model training and testing, we collected, processed, and labeled 551 CT series and 555 MRI series (210 T1 series, 244 T2 series, and 101 FLAIR series). We tested both the CT and MRI models on an unseen dataset containing 83 CT and 84 MRI series, and obtained the following accuracy metrics:

• Average Precision on CT:

  • F1-score = 96.0%

  • AP (Face) = 81.1%

  • AP (Left Ear) = 70.3%

  • AP (Right Ear) = 71.0%

  • mAP = 74.1%

• Average Precision on MRI:

  • F1-score = 93.9%

  • AP (Face) = 97.2%

  • AP (Left Ear) = 92.6%

  • AP (Right Ear) = 88.0%

  • mAP = 92.6%

The performance of the model is as follows:

• GPU (NVIDIA RTX 3090 24 Gb): 37 ms/image

• GPU (AMD Ryzen Threadripper 3960X): 880 ms/image

Collaborators: Fima Furman (Centaur Labs, Boston, United States), Alexey Kachalov (Ambra Health, Boston, United States)

Project type: Commercial

Media: model tests on CT, model tests on MRI