Video-based multi-target multi-camera tracking for postoperative phase recognition.

Area of Science:

• Medical Imaging and Computer Vision
• Surgical Workflow Analysis
• Artificial Intelligence in Healthcare

Background:

• Current deep learning applications for surgical support are primarily focused within the operating room (OR).
• Expanding technological assistance to postoperative workflows presents opportunities for improved patient management and documentation.
• Automating the tracking and phase recognition of patients post-surgery can alleviate manual burdens and enhance data accuracy.

Purpose of the Study:

• To propose and evaluate a novel multi-target multi-camera tracking (MTMCT) architecture.
• To enable automatic recognition of postoperative phases, precise location tracking of patients, and timestamp generation.
• To extend deep learning-based surgical support beyond the OR into postoperative care settings.

Main Methods:

• Development of a custom MTMCT architecture utilizing three RGB cameras.
• Creation of a multi-camera dataset with 19 reenacted postoperative patient flows, including annotated patients and beds.
• Integration of bed and patient tracking per camera, and a patient state module for phase recognition, location, and timestamps.

Main Results:

• The MTMCT architecture demonstrated robust performance in both single- and multi-patient scenarios.
• In multi-patient settings, postoperative phase traversal accuracy reached 84.9 ± 6.0%, with 91.4 ± 1.5% correct timestamp generation.
• Patient tracking IDF1 achieved 92.0 ± 3.6%, with AFLink proving effective for partial trajectory matching.

Conclusions:

• The proposed MTMCT approach shows significant promise for real-time surgeon support.
• This technology lays the groundwork for enhanced clinical documentation in postoperative care.
• The system has the potential to ultimately improve overall patient care and outcomes.