As the use of robotic assisted surgeries (RAS) increases, so does the amount of recording cameras in the operating room (OR). The data from these videos is essential for post-op follow-up, training of surgeons, training of AI, and even legal purposes. However, each procedure creates an enormous amount of data, which requires storage facilities. Additionally, when sharing this data (uploading to cloud, sending to colleagues), all features disclosing identity of the patient or the healthcare provider must be removed.
Removing irrelevant parts of the video
The cameras involved in the OR recording are usually:
- Room Cameras – located in the OR and recording all activities within the OR.
- Endoscope cameras – the “eyes” of the surgeon during the procedure.
- Stereo Cameras – used for 3D perception for the robot or the surgeon.
- Other modalities – ultrasound cameras or X-ray imaging used during the procedure – used for part of the time.
All of the above cameras record continuously, but not all frames are necessary. Room cameras are turned on before the procedure, but preparation of the room for the procedure is not interesting. Endoscope imaging outside the body or blurry images are irrelevant. Ultrasound and X-ray based images (fluoroscopy, angiography, etc.) often have unnecessary frames which can be discarded.
Deep learning algorithms can be trained to detect all the non-interesting data from videos and discard it. Instead of manually deleting parts of the video, the system automatically recognizes which parts of the video are relevant for future use, and then discards the rest. This feature can reduce data volume significantly, thus reducing storage requirements and costs.
As described above, patient privacy has always been a sacred part of healthcare. Surgical visual data should be shared with other clinical centers for training and consultation purposes, without compromising the patient’s or the healthcare provider’s anonymity.
Manually anonymizing surgical videos is a time-consuming process, which is prone to error:
- Erasing the face of the patient or caregiver in all the frames from a procedure – whether from the room camera or possibly when the endoscope was pulled out and exposed a facial feature.
- Deleting each text reference on the image which can identify the patient, caregiver, or even the hospital where the procedure took place.
When examining these manually, it is easy to miss a frame, and identifying a patient in clinical data has dire consequences.
Face detection is a well-established algorithm, which can easily be implemented to recognize all instances where a face is shown in the data from the OR and replace it with a blurred image or another selected anonymization method. Text detection is also easy to implement, and all instances where text appears in the image can be detected and removed as desired.
Adding these features to your device
Any device which extracts video recordings from the OR should have a video cleaner and anonymizer running in the background, to remove the irrelevant data. Current solutions are limited to use in specific modalities or use-cases. Video feeds are more challenging as they require task-specific networks, trained on dedicated data. RSIP Vision has the expertise to provide you with a suitable custom anonymization feature to your device or software. We can achieve accurate results quickly and efficiently, so that your device can reach the market with all the privacy features required for a successful procedure.