TL;DR
This study investigates the effectiveness of three deidentification methods (black box, letterbox, and half-letterbox views) for facial photographs used in publications. The findings reveal that traditional black box methods are surprisingly ineffective, with over half of faces recognized even when heavily obscured. The half-letterbox view emerged as the most effective deidentification strategy, offering a better balance between anonymity and potential clinical utility.
What Problem Does This Paper Address?
Publishing clinical photographs, especially of faces, is crucial for illustrating medical findings, surgical outcomes, and educational content. However, this practice raises significant ethical concerns regarding patient privacy and anonymity. While informed consent is paramount, the increasing trend towards online publishing and social media dissemination means that the potential reach of published images is far greater than in traditional print journals. This necessitates robust deidentification methods to protect patient identities. Existing deidentification techniques, such as simple black boxes over the eyes, have been questioned for their effectiveness and potential to undermine the clinical utility of the image. This paper addresses the critical gap in understanding how effective different deidentification methods truly are in preventing facial recognition, thereby informing better practices for obtaining informed consent.
Methods
The researchers employed a cross-sectional pilot study design using celebrity photographs to simulate real-world scenarios where faces are recognizable.
- Deidentification Techniques: Three methods were tested:
- Black Box: A black box was progressively used to cover facial features, starting with the entire face and gradually revealing more.
- Letterbox View: The image was cropped to reveal only a specific region of the face, progressively revealing more.
- Half Letterbox View: Similar to the letterbox view, but only half of the face was revealed at each stage.
- Image Preparation: 40 well-known celebrity faces were selected. AI-driven facial landmark detection (MT-CNN) was used to identify key points on each face. These landmarks then served as guides for the manual generation of deidentified images in Adobe Photoshop. Seven stages of deidentification were created for each method, culminating in the full, un-deidentified face at stage 8.
- Participant Study: Two surveys were conducted with university students and staff. Participants viewed slideshows of the deidentified celebrity images in sequential stages and were asked to identify the celebrity as soon as they could. Recognition was recorded at the stage it occurred.
Datasets
- Celebrity Image Dataset: 40 well-known celebrity faces were sourced from YouGov's "The most famous people in the UK" list. These images were standardized to 900x900 pixels with a solid black background.
- Deidentified Image Dataset: For each of the 40 celebrity faces, 8 deidentified images were generated for each of the three deidentification methods (black box, letterbox, half-letterbox). This resulted in a total of 480 unique deidentified images used across the study.
Key Insights and Ideas
- Black Boxes are Surprisingly Ineffective: The most striking finding is that even a black box covering the periorbital region, forehead, nose, mouth, and cheeks still allows for over 50% recognition of faces. This challenges the common assumption that such masking is sufficient for anonymity.
- Eyebrows are Key Identifiers: The letterbox view's effectiveness significantly increased once eyebrows were revealed, highlighting their crucial role in facial recognition, as suggested by prior research.
- Holistic Face Recognition: The study implicitly supports the idea that humans recognize faces holistically. When parts of the face are obscured, it cues the viewer to focus on the visible parts and "fill in the blanks," making recognition easier than expected.
- Clinical Utility vs. Anonymity Trade-off: While the half-letterbox view offers the best anonymity, its clinical utility is reduced, particularly for conditions not localized to the visible facial regions. However, it remains valuable for specific specialties like ophthalmology.
Key Results
- Black Box: 55.5% of faces were recognized at the very first stage (black box covering most features). By stage 3 (revealing the nose), recognition reached 94.8%.
- Letterbox View: Recognition was very low initially (2% at stage 1). It surpassed 50% recognition only after stage 4, when eyebrows were included. By stage 6 (revealing the nose), recognition reached 92.5%.
- Half Letterbox View: Recognition was minimal until stage 5. It reached 71.4% at stage 6 (revealing the nose), and 40.9% at stage 5.
- Comparison: The mean stage of recognition was significantly different across methods: Black box (1.7), Letterbox (4.5), and Half Letterbox (5.6). The half-letterbox method was recognized significantly later than the letterbox, and both letterbox methods were substantially superior to the black box.
Limitations and Future Work
- Limitations:
- The study used well-known celebrity images, meaning participants were considering a limited pool of individuals.
- Participants were allowed multiple guesses, which might have influenced recognition rates.
- Manual generation of deidentified images, while guided by AI, could have introduced minor inconsistencies.
- The number of participants was small (n=16 per survey).
- Future Work:
- Further research is needed to test other deidentification methods like blurring, cloning neighboring skin, and pixelation.
- Investigating the acceptance of deidentification methods by patients and stakeholders across different publication modes (online vs. print) is crucial.
- The authors recommend that journals adopt specific consent forms that detail the deidentification method used and confirm that its effectiveness has been discussed with the patient. This information should be integrated into the informed consent process.
Summary updated